A method and system for personalized urban vehicle path planning based on adversarial inverse reinforcement learning

Abstract

This invention provides a personalized route planning method and system for urban vehicles based on adversarial inverse reinforcement learning, belonging to the field of route planning. It addresses the problems of poor adaptability and weak personalization capabilities of existing route planning methods in large-scale, complex urban road networks due to sparse trajectory data. This invention processes historical data and maps it onto the road network, models the environment, facilitates the design of generators and discriminators using convolutional neural networks, constructs learning objectives, and gradually improves the quality of generated data and the discriminator's discriminative ability by alternately optimizing the loss functions of the generator and discriminator. The optimized model is then used for route recommendation. This invention enables the automatic learning of efficient route planning strategies in dynamic traffic environments; simultaneously, by combining historical driving data, it improves the personalized route recommendation capability and the adaptability of the route recommendation system, making the recommended routes more aligned with the actual needs of drivers.

Description

Technical Field

[0001] This invention relates to the field of path planning technology, and more specifically, to a method and system for personalized path planning of urban vehicles based on adversarial inverse reinforcement learning. Background Technology

[0002] With the acceleration of urbanization and the continuous growth of motor vehicle ownership, urban traffic congestion and low travel efficiency are becoming increasingly prominent problems. Traditional vehicle route planning methods mainly rely on static road network data and real-time traffic information (such as congestion index) and use classic algorithms to calculate the shortest or fastest path. However, these methods have the following limitations:

[0003] A Chinese invention patent, CN119417001A, describes a personalized route planning method and system based on federated learning and edge computing. It quantifies a user's travel experience in a visited area into a route trust score (route length, travel time, route safety, and route congestion), and optimizes the route trust score based on the frequency of route selection in the user's social network, providing personalized route planning. Edge computing technology is incorporated to improve the system's efficiency in transmitting and processing data. Another Chinese invention patent, CN119692584A, describes a digital tourist travel planning system that recommends routes based on basic tourist information (personal information, travel preferences, and historical itineraries) and relevant information about target attractions (attraction location, opening hours, ticket prices, and transportation methods). The route planning algorithm is constructed with a set route length as the objective function and commuting time as a constraint, and the planned routes are ranked. However, both of these methods rely on externally specific rules, i.e., setting a uniform weight formula, and do not fully explore the user's potential route selection preferences, resulting in low personalization.

[0004] A Chinese invention patent (patent number: CN115577175A) describes a personalized path recommendation algorithm based on the deep learning-guided A* algorithm. The algorithm involves preprocessing data to extract usable data from users' historical travel information; using a CNN+LSTM neural network to learn the implicit personalized preference features of users from the extracted data; and then using the CNN+LSTM neural network for path planning to obtain candidate paths. The evaluation criteria for path selection serve as the estimation function for the A* search algorithm, used as the basis for selecting the next node in the iteration process of the A* search algorithm. The path planned by the A* search algorithm is the final solution, which is the personalized path recommended to the user. However, the above method only utilizes simulation data and only conducts simple obstacle avoidance experiments on a grid, without using real road data. This results in poor adaptability and inaccurate path planning in large-scale complex urban road networks. Summary of the Invention

[0005] The technical problem to be solved by this invention is:

[0006] To address the problem of poor adaptability and weak personalization capabilities of existing path planning methods in large-scale and complex urban road networks due to sparse trajectory data.

[0007] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0008] This invention provides a personalized route planning method for urban vehicles based on adversarial inverse reinforcement learning, comprising the following steps:

[0009] S100. Data processing: Matching the processed historical trajectory data onto the city road network map; including optimizing the road network structure of the city road network data, processing the historical trajectory data, and then mapping the processed trajectory data onto the road network using a map matching algorithm.

[0010] S200, Model training, including adversarial inverse reinforcement learning, introduces an attention mechanism, and obtains a reward function by maximizing trajectory similarity and minimizing path cost through adversarial training of the generator and discriminator, learns potential path preferences, and captures the multi-objective decision-making strategies of experts through adversarial training and inverse reasoning to generate paths;

[0011] S300 route recommendation provides corresponding route recommendations for both large-scale and small-scale road networks.

[0012] Furthermore, in step S100, the method for processing GPS data includes coordinate system transformation, data filtering, data cleaning, and trajectory segmentation.

[0013] Further, in step S200, the following are included:

[0014] S210. Environmental modeling: Model the road network as a directed graph, denoted by G = (V, E), where V is the set of nodes, representing key locations, and E is the set of edges, representing directed road segments, i.e. road segments where vehicles can travel in a specific direction.

[0015] S220, Generator Design: A generator is designed based on a convolutional neural network. By processing the input state, destination, path features, and environmental features, feasible actions are generated.

[0016] S230, Discriminator Design: The discriminator is designed based on a convolutional neural network. The discriminator extracts neighborhood information from the input features through the convolutional neural network to distinguish the trajectory generated by the generator from the trajectory of a real expert.

[0017] S240, Learning Objectives: Taking into account the path preferences hidden in historical trajectory data and the reward function of travel length, including environmental feature extraction, path cost calculation and design optimization objectives;

[0018] S250, adversarial training, gradually improves the quality of generated data and the discriminator's discrimination ability by alternately optimizing the loss functions of the generator and discriminator, so that the distribution of generated data is as consistent as possible with that of real data.

[0019] Further, in step S220, the convolutional layer processes the input path features. The first convolutional layer uses 20 3*3 convolutional kernels, and the second convolutional layer uses 30 2*2 convolutional kernels to extract local features. The attention mechanism further processes the features output by the convolutional layer to capture the global dependencies between path features. Three fully connected layers are used to extract features step by step to generate action probability distributions. The convolutional layer uses the LeakyReLU activation function, and the pooling layer uses max pooling. The generator output is the probability distribution of each action, obtained through the Softmax function, which is used to select actions or calculate the log probability of actions.

[0020] Further, in step S230, features are first extracted through a convolutional layer, and then the feature map size is reduced and the feature abstraction is increased through a max pooling layer; a second convolutional layer further extracts high-level features, and a multi-attention mechanism module is added after the convolutional layer of the discriminator. The features output by the convolutional layer are processed by the MHSA module, and the resulting feature vector is mapped through a fully connected layer. The Sigmoid activation function is used to calculate the trajectory score, distinguish the difference between the recommended path and the real path, and calculate the path similarity.

[0021] Furthermore, step S240 specifically includes,

[0022] S241. Environmental feature extraction: Extract environmental features related to the trip from historical trajectory data, including road grade, number of turns, and distance to the destination.

[0023] S242. Path cost calculation: Define path cost as travel distance, extract road length information from the road network, and calculate the cost of each path.

[0024] S243. Optimize the objective by integrating environmental characteristics and path costs. The optimization objective takes the following form:

[0025] R(s,a)=R expert (s,a)+αR cost (s,a)

[0026] Among them, R expert (s,a) represents the reward for learning based on environmental features, Rcost (s,a) represents the path cost reward, and α represents the hyperparameter, which adjusts the weight between the two rewards.

[0027] Further, in step S250, an alternating training method is adopted. First, the discriminator D is fixed, and the loss function of the generator G is optimized. The optimization objective of the generator is to maximize the negative logarithm of the probability that generated data is classified as real data in the discriminator. Then, the generator G is fixed, and the loss function of the discriminator D is optimized. The optimization objective of the discriminator is to maximize the ability to classify real data as real data. The optimization objectives of the generator and the discriminator are as follows:

[0028]

[0029] Where z represents the input to the generator, G(z) represents the data generated by the generator, D(·) represents the discriminator function, which can represent the probability that the input data is real data; x represents the collected real-world data, p data (x) represents the distribution of the real data, and p(z) represents the distribution of the data generated by the generator.

[0030] Furthermore, step S300 specifically includes,

[0031] S310. For small-scale road networks, where the trajectory distribution is relatively dense, the action selection probability output by the model is used directly for path finding.

[0032] S320. For large-scale road networks, since the trajectory distribution is relatively sparse, the model is first used to obtain the selection probability of each action in the current state. The reciprocal of the action selection probability is used as the cost of Dijkstra's algorithm for path finding. The optimized model can select the overall optimal path based on the global selection.

[0033] A personalized urban vehicle path planning system based on adversarial inverse reinforcement learning is provided. The system has program modules corresponding to the steps described above, and executes the steps in the personalized urban vehicle path planning method based on adversarial inverse reinforcement learning during runtime.

[0034] A computer-readable storage medium storing a computer program configured to, when invoked by a processor, implement steps of a personalized urban vehicle path planning method based on adversarial inverse reinforcement learning.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] This invention presents a personalized route planning method and system for urban vehicles based on adversarial inverse reinforcement learning. Compared with traditional route recommendation algorithms, such as discrete choice models or experience-based shortest path algorithms, it exhibits stronger adaptability and generalization ability. Through adversarial training, this invention can not only learn the optimal route selection from expert examples (historical trajectories), but also accurately capture users' personalized travel preferences, overcoming the shortcomings of traditional algorithms that cannot flexibly adapt to individual needs.

[0037] Traditional route planning algorithms, particularly shortest path algorithms (such as A* and Dijkstra's algorithms), focus on optimizing path length while neglecting personalized user needs, such as driving habits and preferred route types (e.g., avoiding congested areas or choosing expressways). Machine learning-based route recommendation methods are widely used in multi-objective route optimization problems. However, the objectives of route planning are also human-designed, and different travelers have different perceptions of road networks and cognitive biases in understanding the complex factors influencing route selection. Even with the same starting point and destination, individuals may have different route preferences. Therefore, the recommendation results of traditional route selection models are difficult to fully meet the actual needs of users.

[0038] Existing path planning models based on adversarial inverse reinforcement learning can learn potential path selection preferences from expert examples (historical trajectories), but they do not take into account the impact of the actual travel environment, are highly dependent on data, and are prone to getting stuck in local optimization during the path finding process. In large-scale or sparse road networks, the path recommendation results are poor.

[0039] This invention improves the generator and discriminator of generative adversarial networks (GANs). The generator considers the influence of the environment on path selection when learning historical trajectories, while the discriminator considers the influence of both path preference and path length on path selection. Specifically, when imitating expert strategies, an additional path length reward is introduced. By introducing this reward, the discriminator can guide the trajectory closer to the target location, rather than simply imitating expert data, reducing local optima and improving overall path quality. Secondly, an attention layer is added after the convolutional neural networks of the generator and discriminator, strengthening the information flow during path selection and better addressing the dynamically changing traffic environment. Furthermore, this invention optimizes the path generation strategy by using the output of a path planning model based on adversarial inverse reinforcement learning as the weight for Dijkstra's algorithm. This reduces the model's dependence on trajectory data, preserving the model's trajectory selection preferences while effectively controlling path length, making the trajectory structure closer to real data, and achieving a fusion of model preferences and global optimization.

[0040] Experimental results show that this invention can combine the model's performance in reducing path length and approximating user travel preferences. With fewer training iterations, it can achieve better path recommendation results, that is, it has achieved good results in both fitting user travel preferences and reducing path length, improving the ability of personalized path recommendation and the adaptability of the path recommendation system, making the recommended paths more in line with the actual needs of drivers; and this invention is applicable to complex small-scale cities and large-scale cities. Attached Figure Description

[0041] Figure 1 This is a flowchart of a personalized urban vehicle path planning method based on adversarial inverse reinforcement learning, as described in an embodiment of the present invention.

[0042] Figure 2 This is a flowchart of a path recommendation process for a personalized urban vehicle path planning method based on adversarial inverse reinforcement learning, according to an embodiment of the present invention.

[0043] Figure 3 This is a structural diagram of the generator in an embodiment of the present invention;

[0044] Figure 4 This is a structural diagram of the discriminator in an embodiment of the present invention;

[0045] Figure 5 This refers to the road network within the Shanghai dataset in this embodiment of the invention;

[0046] Figure 6 This refers to the road network within the Wuhan city dataset in this embodiment of the invention.

[0047] Figure 7 This is a matching map of real historical GPS trajectories on the road network of Wuhan City in an embodiment of the present invention;

[0048] Figure 8 This is a result diagram of route recommendations in Wuhan City in an embodiment of the present invention. Detailed Implementation

[0049] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0050] Specific Implementation Plan 1: Combining Figures 1 to 4 As shown, this invention provides a personalized urban vehicle path planning method based on adversarial inverse reinforcement learning, comprising the following steps:

[0051] S100. Data processing: Optimize the road network structure of urban road network data, such as performing coordinate system transformation, data filtering, data cleaning, and trajectory segmentation on taxi GPS data. Then, use map matching algorithms to accurately map the processed trajectory data onto the road network using fastmap matching (FMM). Figure 5 The matching results shown;

[0052] S200, Model Training, including research on path selection methods based on adversarial inverse reinforcement learning and the introduction of attention mechanisms. Through adversarial training of the generator and discriminator, a robust reward function is obtained with the objectives of maximizing trajectory similarity and minimizing path cost. Potential path preferences are learned, and through adversarial training and inverse reasoning, multi-objective decision-making strategies of experts are captured to generate paths that meet various needs; specifically including…

[0053] S210. Environmental Modeling: The road network is modeled as a directed graph, denoted by G = (V, E), where V is the set of nodes, representing key locations (such as intersections or the start and end points of roads), and E is the set of edges, representing directed road segments, i.e., road segments that vehicles can travel in a specific direction. The connectivity of road segments is determined by their spatial location, i.e., one road segment can connect to another road segment to form a network. Each road has a feature vector, mainly considering road grade, number of turns, and distance to the destination. These features can be mined from historical GPS trajectory data to provide references or recommendations for similar travel needs.

[0054] S220. Generator Design: This invention designs a generator based on a Convolutional Neural Network (CNN). By processing input state, destination, path features, and environmental features, it generates feasible actions. Convolutional layers (conv1, pool, conv2) first process the input path features. The first convolutional layer uses 20 3x3 kernels, and the second convolutional layer uses 30 2x2 kernels to extract local features. An attention mechanism further processes the features output by the convolutional layers, capturing the global dependencies between path features. Three fully connected layers are used to progressively extract features and generate action probability distributions. The convolutional layers use the LeakyReLU activation function, and the pooling layers use max pooling. The generator's output is the probability distribution of each action, obtained through the Softmax function. These probability distributions are used to select actions or calculate the log probability of actions. The generator structure is as follows: Figure 3 As shown;

[0055] S230. Discriminator Design: This invention designs a discriminator based on a Convolutional Neural Network (CNN) to distinguish between trajectories generated by the generator and real expert trajectories. The discriminator extracts neighborhood information from the input features through the CNN. First, features are extracted through a convolutional layer, followed by a max-pooling layer to reduce the feature map size and increase the level of feature abstraction. A second convolutional layer further extracts high-level features. A multi-attention mechanism module is added after the convolutional layers of the discriminator. The features output from the convolutional layers are processed by the MHSA module, and the resulting feature vector is mapped through a fully connected layer. The Sigmoid activation function is used to calculate the trajectory score, distinguish the difference between the recommended path and the real path, and calculate the path similarity. The discriminator structure is as follows: Figure 4 As shown;

[0056] S240. Learning Objective: This invention can learn a reward function that comprehensively considers the path preferences and trip length hidden in historical trajectory data, so that the generated path not only conforms to the driver's path preferences, but also finds the optimal balance among multiple objectives; specifically including,

[0057] S241. Environmental feature extraction: Extracting environmental features related to the trip from historical trajectory data, including road grade, number of turns, and distance to the destination; the environmental feature matrix is ​​a three-dimensional matrix that comprehensively describes the complexity and cost of the path and is crucial for path selection.

[0058] S242. Path cost calculation: Define path cost as travel distance, extract road length information from the road network, and calculate the cost of each path.

[0059] S243. Optimize the objective by integrating environmental characteristics and path costs. The optimization objective takes the following form:

[0060] R(s,a)=R expert (s,a)+αR cost (s,a)

[0061] Among them, R expert (s,a) represents the reward for learning based on environmental features, reflecting the model's effectiveness in fitting path preferences; R cost (s,a) represents the path cost reward, and α represents the hyperparameter, which adjusts the weight between the two rewards. In this invention, path cost and path preference are considered comprehensively. While not affecting the model's ability to learn from environmental features and fit path preferences, the path cost is reduced as much as possible. Therefore, the value of α is set to 1.5.

[0062] S250. Adversarial training: The model training adopts an alternating training method. First, the discriminator D is fixed, and the loss function of the generator G is optimized. The optimization objective of the generator is to maximize the negative logarithm of the probability that generated data is classified as real data in the discriminator. Then, the generator G is fixed, and the loss function of the discriminator D is optimized. The optimization objective of the discriminator is to maximize its ability to classify real data as real data. The optimization objectives of the generator and the discriminator are as follows:

[0063]

[0064] Where z represents the input to the generator, G(z) represents the data generated by the generator, D(·) represents the discriminator function, which can represent the probability that the input data is real data; x represents the collected real-world data, p data (x) represents the distribution of the real data, and p(z) represents the distribution of the data generated by the generator;

[0065] By alternately optimizing the loss functions of the generator and discriminator, the quality of the generated data and the discriminator's discrimination ability are gradually improved, so that the distribution of the generated data is as consistent as possible with that of the real data.

[0066] S300, route recommendation, including route recommendation algorithms for large-scale and small-scale road networks;

[0067] S310. For small-scale road networks, where the trajectory distribution is relatively dense, the action selection probability output by the model is used directly for path finding.

[0068] S320. For large-scale road networks, since the trajectory distribution is relatively sparse, the "reciprocal of the action probability" output by the model is defined as the path cost of Dijkstra's algorithm. A higher probability action means a lower cost path, thereby reducing the model's dependence on data coverage. First, the model is used to obtain the selection probability of each action in the current state. The reciprocal of the action selection probability is used as the path finding cost of Dijkstra's algorithm. The optimized model can select the overall optimal path based on the global data, which not only retains the model's trajectory selection preference, but also effectively controls the path length, making the trajectory structure closer to the real data, and realizing the fusion of model preference and global optimization.

[0069] Specific Implementation Scheme 2: This invention provides a personalized urban vehicle path planning system based on adversarial inverse reinforcement learning. The system has program modules corresponding to the above steps, and executes the steps in the above-mentioned personalized urban vehicle path planning method based on adversarial inverse reinforcement learning when running.

[0070] The other combinations and connections in this implementation scheme are the same as in Specific Implementation Scheme 1.

[0071] Specific Implementation Scheme 3: The present invention provides a computer-readable storage medium storing a computer program configured to implement, when called by a processor, a method for personalized urban vehicle path planning based on adversarial inverse reinforcement learning.

[0072] The other combinations and connections in this implementation scheme are the same as in Specific Implementation Scheme 1.

[0073] experiment

[0074] The effectiveness of the proposed model was validated using a small-scale Shanghai dataset and a large-scale Wuhan dataset. The road network within the Shanghai dataset is as follows: Figure 5 As shown, the road network within the Wuhan city dataset is as follows: Figure 6 As shown in Table 1, the experimental dataset is presented.

[0075] Table 1. Dataset Description

[0076]

[0077]

[0078] Evaluation metrics: 1. Fitting path preference: Trajectory structure similarity ED (the smaller the better), node matching degree BLUE (the larger the better), distribution fit degree JSD (the smaller the better) 2. Reduce path length: Average path length (the smaller the better).

[0079] The performance of MR-AIRL of this invention compared with other path recommendation algorithms in terms of ED, BLUE, JSD, and average path length is shown in Tables 2 to 5 below. Since the average values ​​of the metrics in the experiments in Shanghai and Wuhan more clearly reflect the experimental results of the model, the specific values ​​of each metric across the eight sets of experiments are not listed here. The average values ​​of each metric in the experiments in Shanghai and Wuhan are calculated separately to analyze the overall performance of each path recommendation method. The overall performance of MR-AIRL of this invention compared with other path recommendation algorithms on eight datasets in the two cities across the four performance metrics is shown in Tables 6 and 7 below.

[0080] Table 2 Comparison of results from different path recommendation methods (ED)

[0081]

[0082] Table 3 Comparison of BLEU results for different path recommendation methods.

[0083]

[0084]

[0085] Table 4 Comparison of results from different path recommendation methods using JSD

[0086]

[0087] Table 5 Comparison of average path length results for different path recommendation methods

[0088]

[0089] Table 6 shows the overall performance in Shanghai.

[0090]

[0091] Table 7 Overall Performance in Wuhan

[0092]

[0093] Experimental Results: The model designed in this invention comprehensively demonstrates its effectiveness in both reducing path length and closely approximating user travel preferences. Experimental results on the Shanghai dataset show that the MR-AIRL model performs best across all four metrics, with significant improvements in Path Structure Consistency (ED) and Node Matching (BLEU). The average path length is also close to that of the A* algorithm, resulting in optimal overall performance. The Shanghai dataset shows that the model outperforms Dijkstra's algorithm by 82.2%, 341%, and 72% in trajectory structure, node matching, and distribution fitting, respectively, reducing the average recommended path length by 17.8%. Compared to the A* algorithm, it outperforms the A* algorithm by 71.7%, 160%, and 54.3% in these areas, achieving better path recommendation results with fewer training iterations. Compared to the AIRL model that only considers path preferences, the model designed in this invention improves upon these metrics by 48.4%, 26.5%, and 20.1%, respectively, reducing the average recommended path length by 7.8%.

[0094] The matching map of the actual historical trajectory on the road network of Wuhan is as follows: Figure 7 As shown, the results of path recommendation using the model of this invention are as follows: Figure 8As shown in the figure. Experimental results on the Wuhan dataset show that the model outperforms Dijkstra's algorithm by 83.74%, 184.04%, and 62.02% in trajectory structure, node matching, and distribution fitting, respectively, and reduces the average recommended path length by 13.97%. Compared to the A* algorithm, it outperforms it by 73.05%, 77.75%, and 55.03% in trajectory structure, node matching, and distribution fitting, respectively, achieving better path recommendation results with fewer training iterations. Compared to the AIRL model that only considers path preferences, the model designed in this invention outperforms Dijkstra's algorithm by 78.6%, 102.5%, and 46.9% in trajectory structure, node matching, and distribution fitting, respectively, and reduces the average recommended path length by 62.6%.

[0095] Based on the two experiments above, the model of this invention outperforms other models on both small-scale road networks in Shanghai and large-scale road networks in Wuhan, especially in the case of route recommendation problems with large-scale road networks and sparse data, where it demonstrates stronger adaptability.

[0096] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A personalized route planning method for urban vehicles based on adversarial inverse reinforcement learning, characterized in that, Includes the following steps: S100. Data processing: Matching the processed historical trajectory data onto the city road network map; including optimizing the road network structure of the city road network data, processing the historical trajectory data, and then mapping the processed trajectory data onto the road network using a map matching algorithm. S200, Model training, including adversarial inverse reinforcement learning, introduces an attention mechanism, and obtains a reward function by maximizing trajectory similarity and minimizing path cost through adversarial training of the generator and discriminator, learns potential path preferences, and captures the multi-objective decision-making strategies of experts through adversarial training and inverse reasoning to generate paths; include, S210, Environmental Modeling: Model the road network as a directed graph, using... It means that, among them, It is a set of nodes, representing key positions. It is a set of edges, representing directed road segments, that is, road segments where vehicles can travel in a specific direction; S220, Generator Design: A generator is designed based on a convolutional neural network. By processing the input state, destination, path features, and environmental features, feasible actions are generated. S230, Discriminator Design: The discriminator is designed based on a convolutional neural network. The discriminator extracts neighborhood information from the input features through the convolutional neural network to distinguish the trajectory generated by the generator from the trajectory of a real expert. include, First, features are extracted through a convolutional layer, followed by a max pooling layer to reduce the feature map size and increase the level of feature abstraction. A second convolutional layer further extracts high-level features. After the convolutional layer of the discriminator, a multi-attention mechanism module is added. The features output by the convolutional layer are processed by the MHSA module, and the resulting feature vector is mapped through a fully connected layer. The Sigmoid activation function is used to calculate the trajectory score, distinguish the difference between the recommended path and the real path, and calculate the path similarity. S240, Learning Objectives: Taking into account the path preferences hidden in historical trajectory data and the reward function of travel length, including environmental feature extraction, path cost calculation and design optimization objectives; include, S241. Environmental feature extraction: Extract environmental features related to the trip from historical trajectory data, including road grade, number of turns, and distance to the destination. S242. Path cost calculation: Define path cost as travel distance, extract road length information from the road network, and calculate the cost of each path. S243. Optimize the objective by integrating environmental characteristics and path costs. The optimization objective takes the following form: in, Rewards representing learning based on environmental characteristics. Represents path cost reward, This represents a hyperparameter that adjusts the weight between two rewards. S250, adversarial training, gradually improves the quality of generated data and the discriminator's discrimination ability by alternately optimizing the loss functions of the generator and discriminator, so that the distribution of generated data is as consistent as possible with that of real data; S300 route recommendation provides corresponding route recommendations for both large-scale and small-scale road networks.

2. The method for personalized urban vehicle path planning based on adversarial inverse reinforcement learning according to claim 1, characterized in that: In step S100, the methods for processing GPS data include coordinate system transformation, data filtering, data cleaning, and trajectory segmentation.

3. The method for personalized urban vehicle path planning based on adversarial inverse reinforcement learning according to claim 1, characterized in that: In step S220, the convolutional layer processes the input path features. The first convolutional layer uses 20 3*3 convolutional kernels, and the second convolutional layer uses 30 2*2 convolutional kernels to extract local features. The attention mechanism further processes the features output by the convolutional layer to capture the global dependencies between path features; three fully connected layers are used to extract features step by step to generate action probability distributions; the convolutional layer uses the LeakyReLU activation function, and the pooling layer uses max pooling; the generator output is the probability distribution of each action, obtained through the Softmax function, which is used to select actions or calculate the log probability of actions.

4. The method for personalized urban vehicle path planning based on adversarial inverse reinforcement learning according to claim 1, characterized in that: In step S250, an alternating training method is adopted. First, the discriminator D is fixed, and the loss function of the generator G is optimized. The optimization objective of the generator is to maximize the negative logarithm of the probability that generated data is classified as real data in the discriminator. Then, the generator G is fixed, and the loss function of the discriminator D is optimized. The optimization objective of the discriminator is to maximize the ability to classify real data as real data. The optimization objectives of the generator and the discriminator are as follows: min min in, Represents the input of the generator. This represents the data generated by the generator. This represents the discriminator function, which indicates the probability that the input data is real data. Represents the collected real-world data, The distribution representing real data, This represents the distribution of data generated by the generator.

5. The method for personalized urban vehicle path planning based on adversarial inverse reinforcement learning according to claim 1, characterized in that: Step S300 specifically includes, S310. For small-scale road networks, where the trajectory distribution is relatively dense, the action selection probability output by the model is used directly for path finding. S320. For large-scale road networks, since the trajectory distribution is relatively sparse, the model is first used to obtain the selection probability of each action in the current state. The reciprocal of the action selection probability is used as the cost of Dijkstra's algorithm for path finding. The optimized model can select the overall optimal path based on the global selection.

6. A personalized route planning system for urban vehicles based on adversarial inverse reinforcement learning, characterized in that: The system has a program module corresponding to the steps of any one of the claims 1-5 above, and executes the steps in the above-described method for personalized urban vehicle path planning based on adversarial inverse reinforcement learning when it is run.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program configured to, when invoked by a processor, implement the steps of any one of claims 1-5: a method for personalized urban vehicle path planning based on adversarial inverse reinforcement learning.

Images