An AI-based method and system for optimizing autonomous driving mobility services

Abstract

This invention discloses an artificial intelligence-based method and system for optimizing autonomous driving travel services, relating to the field of intelligent transportation technology. It includes: S1: Extracting corresponding road network spatiotemporal features and multi-dimensional passenger embedding vectors, and obtaining the corresponding optimal dispatch strategy and carpooling route planning scheme; S2: Obtaining the corresponding comfort loss value through a generative adversarial network model, and correcting the reward function to obtain the corrected optimal dispatch strategy and carpooling route planning scheme; S3: When multiple passengers share a ride, determining the corresponding fare discount ratio through the Shapley algorithm and attention mechanism, and determining the corresponding compensation scheme based on the comfort loss value and the expected compensation amount when the travel time is extended. This invention not only ensures the overall operational efficiency of the fleet but also proactively avoids behaviors that damage user expectations, such as excessive detours and severe time delays, thereby reducing the risk of user complaints and churn from the root.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, specifically to an artificial intelligence-based method and system for optimizing autonomous driving travel services. Background Technology

[0002] Currently, autonomous driving services primarily rely on dedicated vehicles equipped with sensors such as LiDAR and cameras to collect data, which is then combined with high-precision maps for localization and navigation. By identifying stationary structures like traffic lights and guardrails, as well as moving targets like pedestrians and vehicles, a model of the surrounding environment is constructed. However, the production and maintenance of high-precision maps are extremely costly, and they often suffer from update delays, making it difficult to reflect dynamic changes such as road construction and temporary detours in real time. Furthermore, relying solely on onboard sensors significantly reduces perception capabilities in adverse weather conditions (such as rain, snow, and fog) or complex lighting conditions, making it difficult for the system to accurately identify environmental information and impacting the safety and continuity of travel services.

[0003] Chinese invention patent CN113537445A discloses a trajectory prediction method, apparatus, device, and storage medium. The method includes: determining a first object whose trajectory to be predicted in a current autonomous driving scenario; determining a second object associated with the first object based on current object information corresponding to each object in the current autonomous driving scenario; and predicting the target future trajectory of the first object based on a preset trajectory generation model, first historical trajectory information of the first object, and second historical trajectory information of the second object. The technical solution of this invention can improve trajectory prediction efficiency while ensuring accuracy.

[0004] However, the above-mentioned and similar technical solutions still have the following shortcomings: In the shared mobility scenario, in order to reduce the empty driving rate and improve vehicle utilization, the system often forcibly introduces detour pick-up or multi-point pick-up and drop-off mechanisms when dispatching orders, but ignores the differentiated needs of different passengers in terms of time window constraints, comfort expectations and willingness to transfer, which leads to a significant increase in the actual travel time of some passengers, uncontrollable waiting time, and even situations where excessive detours violate the user's initial travel expectations. This will lead to user dissatisfaction, increased complaint rate and decreased platform repurchase rate, thereby limiting the sustainable promotion and commercialization of autonomous driving mobility services in large-scale shared scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence-based method and system for optimizing autonomous driving travel services, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for optimizing autonomous driving travel services based on artificial intelligence, comprising: S1: Multi-objective planning of data fusion: Through spatiotemporal graph attention network and hybrid feature extraction model, the corresponding spatiotemporal features of road network and multidimensional passenger embedding vector are extracted. At the same time, the global scheduling model is trained through the spatiotemporal features of road network and multidimensional passenger embedding vector. After training, the global scheduling model outputs the corresponding optimal dispatch strategy and carpooling route planning scheme. S2: Dynamic correction of comfort perception: The corresponding comfort loss value is obtained through the generative adversarial network model, and the reward function of the trained global scheduling model is corrected according to the comfort loss value to obtain the corrected optimal dispatch strategy and carpooling path planning scheme. S3: Fair Game Benefit Compensation: When multiple passengers share a ride, the Shapley value algorithm and attention mechanism are used to determine the fare discount ratio for each passenger. When the journey time is extended, the corresponding compensation plan is determined based on the aforementioned comfort loss value and the expected compensation amount.

[0007] Furthermore, the output retrieves the corresponding optimal dispatch strategy and carpooling route planning scheme, including: S1.1: Perception Fusion: Through multiple perception channels set up in parallel, corresponding external environment data and travel user data are collected and acquired. Then, through a spatiotemporal graph attention network and a hybrid feature extraction model, features are extracted from the external environment data and travel user data respectively. S1.2: Function Construction: Each autonomous vehicle in the autonomous vehicle fleet is set as a corresponding intelligent agent. Based on the current location, remaining battery power, passenger information, and current time of the autonomous vehicle, the state space of the intelligent agent is set. Based on the order action, route selection, and service strategy of the autonomous vehicle, the action space of the intelligent agent is set. At the same time, based on the state space and action space, the corresponding reward function is constructed. S1.3: Function Evaluation: Based on the road network spatiotemporal feature matrix and passenger embedding vector, a corresponding candidate solution set is constructed, and the score reward value corresponding to each candidate solution in the candidate solution set is obtained through the reward function. At the same time, the corresponding optimal candidate solution is determined through the meta-reinforcement learning algorithm, including the optimal dispatch strategy and the carpooling route planning scheme.

[0008] Furthermore, feature extraction is performed on both external environment data and travel user data, including: S1.1.1: Feature Extraction: Based on the intersections and road segments of the urban road traffic system, road segment nodes and road segment connection edges with weighted directed graph structures are set to construct the corresponding urban road network dynamic graph. At the same time, through the urban road network dynamic graph, the corresponding spatial association weights and temporal association weights are obtained, and the spatial association weights and temporal association weights are fused to obtain the corresponding spatiotemporal joint features. The spatiotemporal joint features are combined with the road segment nodes to construct the corresponding road network spatiotemporal feature matrix. S1.1.2: Vector Generation: By extracting the corresponding trip performance, service feedback, behavioral stability and travel patterns from the passenger IDs corresponding to historical order data, and by extracting the corresponding real-time behavioral data from real-time new orders, the corresponding temporal features are extracted based on the real-time behavioral data and historical order data. The complete behavioral distribution of passengers is compressed and represented by a variational autoencoder to extract the corresponding implicit preferences. The temporal features and implicit preferences are then concatenated to construct the corresponding passenger embedding vector.

[0009] Furthermore, based on the state space and action space corresponding to the agent, the system efficiency reward, detour penalty, expected violation penalty, and churn risk penalty corresponding to the agent are determined. At the same time, based on the dynamic map of the urban road network, the region label corresponding to the agent is determined. The region label and scene label are combined to determine the weight combination corresponding to the agent. The weight combination is then combined with the system efficiency reward, detour penalty, expected violation penalty, and churn risk penalty to obtain the corresponding reward function.

[0010] Furthermore, the revised optimal dispatch strategy and ride-sharing route planning scheme are obtained, including: S2.1: Loss Calculation: Based on the road condition parameters, vehicle dynamics parameters and path geometry information corresponding to each candidate path in the candidate solution set, set the input of the generator in the generative adversarial network model and output the corresponding three-dimensional acceleration change sequence. At the same time, based on the passenger's historical physiological feedback data, perform adversarial training on the discriminator in the generative adversarial network model and output the corresponding comfort loss value and psychological expected compensation amount range. S2.2: Function Compensation: Based on the comfort loss value and the expected compensation amount range, the reward function is modified to obtain the corresponding final modified reward function; S2.3: Feedback Correction: Based on the final corrected reward function, feedback step S1.2 is performed to correct the reward function, and based on the corrected reward function, step S1.3 is executed to determine the corresponding updated optimal candidate solution.

[0011] Furthermore, based on the aforementioned comfort loss value, the reward function is initially modified to obtain an initially modified reward function, specifically as follows: ; in: This is the initially revised reward function. Assigning a weight to the comfort penalty This represents the loss of comfort. For the reward function; Based on the maximum and minimum compensation amounts within the range of expected psychological compensation amounts, the corresponding average compensation amount is determined, and a corresponding expected psychological compensation amount is set. Simultaneously, based on the expected psychological compensation amount, the initially revised reward function is modified to obtain the corresponding final revised reward function, specifically as follows: ; in: For the final revised reward function, This is the initially revised reward function. To compensate for cost weighting, This refers to the amount of compensation expected by the individual.

[0012] Furthermore, the corresponding compensation plan was determined, including: S3.1: Contribution Calculation: Based on the updated optimal dispatch strategy and the updated ride-sharing route planning scheme for each passenger, determine the marginal contribution of each passenger. At the same time, based on the detour mileage and time delay for each passenger, determine the comprehensive contribution of each passenger. Combine the comprehensive contribution with the corresponding dominance weight to determine the discount weight for each passenger. S3.2: Dynamic Compensation: By setting time trigger conditions and comfort trigger conditions, passengers who need dynamic compensation are identified, and based on the range of expected compensation amounts for the corresponding passengers, and based on the passenger embedding vector, passenger levels are divided, and the range of expected compensation amounts and passenger levels are combined to determine the corresponding final compensation amount.

[0013] Furthermore, the discount weight for each passenger is determined, including: S3.1.1: Contribution Calculation: Based on the arrival time and individual travel time corresponding to the updated carpooling route planning scheme, the travel time difference for each passenger is determined. Then, using the Shapley value algorithm, the updated carpooling route planning scheme, the travel time difference, and the passenger embedding vector are combined to obtain the marginal contribution for each passenger. Based on this marginal contribution, the comprehensive contribution of each passenger is determined, specifically as follows: ; in: The overall contribution of the p-th passenger. This is the mileage weighting coefficient. For time weighting coefficients, Let p be the time sensitivity weight of the p-th passenger. The extra distance traveled by passengers The time delay caused by the p-th passenger. The time delays caused to passengers The detour distance caused by the p-th passenger; S3.1.2: Dominance Identification: The overall contribution, time delay, time sensitivity weight, comfort tolerance threshold, and historical churn probability are concatenated and combined to construct a corresponding co-multiplicative feature vector. Then, through an attention mechanism, the dominance weight corresponding to each passenger is obtained, specifically: ; in: Let p be the weight of the degree of dominance of the p-th passenger. Let be the shared ride feature vector of the p-th passenger. Let v be the ride-sharing feature vector of the v-th passenger. It is an exponential function. , Indexing for passengers, Gather passengers for the current journey; S3.1.3: Discount Determination: The final discount weight for the passenger is determined by combining the overall contribution and dominance weights, specifically as follows: ; in: Let the final discount weight be for the p-th passenger. The dominant amplification factor (e.g., 0.5-1). Let p be the weight of the degree of dominance of the p-th passenger. The overall contribution of the p-th passenger.

[0014] Furthermore, the passenger level includes high level, medium level and low level. When the passenger level is high level, the corresponding final compensation amount is set according to the maximum compensation amount in the range of expected compensation amount. When the passenger level is low, the corresponding final compensation amount is set according to the minimum compensation amount in the range of expected compensation amount. When the passenger level is medium, the maximum and minimum compensation amounts within the expected compensation range are combined with a compensation bias coefficient to set the corresponding final compensation amount, specifically: ; in: The final compensation amount, To compensate for the bias coefficient, The maximum compensation amount, This is the minimum compensation amount.

[0015] An AI-based autonomous driving mobility service optimization system uses any one of the aforementioned AI-based autonomous driving mobility service optimization methods.

[0016] Compared with the prior art, the beneficial effects of the present invention are: Firstly, this invention integrates the spatiotemporal dynamic features of the road network through a spatiotemporal graph attention network and combines it with multi-dimensional passenger embedding vectors that cover users' historical behavior, real-time preferences, and implicit needs. This breaks through the limitations of traditional scheduling that only focuses on vehicle empty-running rate and utilization rate. At the same time, it constructs a multi-dimensional reward function that includes system efficiency, detour penalty, expectation violation penalty, and churn risk penalty. The weights can be dynamically adjusted according to region and scenario. This not only ensures the overall operational efficiency of the fleet but also avoids behaviors that damage user expectations, such as excessive detours and serious timeouts, in advance, thereby reducing user complaints and churn risk from the root. Secondly, this invention uses generative adversarial networks combined with historical physiological feedback data from passengers to quantify the loss of comfort caused by changes in travel acceleration and incorporate it into the correction term of the scheduling reward function. It also links it to the range of psychological expectation compensation amount, so that scheduling decisions can proactively avoid highly uncomfortable paths such as sudden acceleration and deceleration and frequent lane changes. This solves the problem of poor riding experience caused by the "efficiency-only" approach of traditional solutions, and thus adapts to the needs of passengers who are sensitive to bumps and sudden braking. Thirdly, this invention quantifies the actual contribution of each passenger to detours and delays through the Shapley value algorithm, and combines it with an attention mechanism to identify the dominant passenger with a higher influence on the trip decision, so as to dynamically adjust the discount weight. This can avoid the problem of "unequal responsibility and unfair sharing", so that the carpooling cost can match the actual travel benefits obtained by the passenger. Fourthly, this invention uses time and comfort as dual triggering conditions to match different levels of passengers with differentiated compensation amounts. That is, high-value users receive full compensation and price-sensitive users receive basic protection. This not only controls the platform's compensation costs but also accurately soothes the dissatisfaction of different groups, effectively improving users' acceptance of autonomous driving shared mobility and their willingness to repurchase. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the autonomous driving mobility service optimization method of the present invention; Figure 2This is a flowchart illustrating the dynamic correction method for comfort perception in this invention. Figure 3 This is a schematic diagram of the correction process for the reward function in this invention; Figure 4 This is a schematic diagram illustrating the process of determining the final discount weight in this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] refer to Figure 1 This embodiment provides an artificial intelligence-based method for optimizing autonomous driving mobility services, which specifically includes the following steps: Step S1: Multi-objective planning for data fusion. This involves processing the road network topology data using a spatiotemporal graph attention network to extract the corresponding spatiotemporal features of the road network. Simultaneously, historical order data and real-time behavior data of passengers are collected, and corresponding multi-dimensional passenger embedding vectors are extracted using a pre-defined hybrid feature extraction model.

[0020] Furthermore, a multi-agent Markov decision process is used to define the scheduling strategy for autonomous vehicles, constructing a corresponding global scheduling model. Simultaneously, the constructed global scheduling model is trained using extracted road network spatiotemporal features and multi-dimensional passenger embedding vectors, resulting in a post-trained global scheduling model obtained through a meta-reinforcement learning algorithm. Finally, the trained global scheduling model outputs the optimal dispatch strategy and carpooling route planning scheme.

[0021] Step S2: Dynamic Correction of Comfort Perception. This involves simulating corresponding driving experience data using a generative adversarial network model to obtain the corresponding comfort loss value. Based on this comfort loss value, the reward function of the global scheduling model trained in Step S1 is corrected, thereby refining the optimal dispatch strategy and carpooling route planning scheme obtained in Step S1.

[0022] Step S3: Fair Game-Based Benefit Compensation. When multiple passengers share a ride, the Shapley value algorithm quantifies each passenger's contribution to the overall detour, and an attention mechanism identifies the dominant passenger to dynamically adjust each passenger's fare discount ratio. Simultaneously, when passenger travel time is extended due to carpooling, the comfort loss value obtained in Step S2 is combined with the expected compensation amount to generate a corresponding compensation plan.

[0023] This embodiment also provides an artificial intelligence-based autonomous driving travel service optimization system, which uses the aforementioned artificial intelligence-based autonomous driving travel service optimization method.

[0024] In this embodiment, the global scheduling model corresponding to the scheduling strategy of autonomous vehicles is trained by extracting the spatiotemporal features of the road network and multi-dimensional passenger embedding vectors. The trained global scheduling model then outputs the corresponding optimal dispatch strategy and ride-sharing route planning scheme. (Reference) Figure 2 This embodiment provides a multi-objective planning method for data fusion, which specifically includes the following steps: Step S1.1: Perceptual Fusion. This involves acquiring corresponding external environment data and travel user data through two parallelly configured perceptual channels. The acquired external environment data is dynamically modeled using a spatiotemporal graph attention network, and the acquired travel user data is fused and represented using a variational autoencoder and a temporal feature extraction model. Details are as follows: Step S1.1.1: Feature Extraction. This involves setting up road segment nodes in a weighted directed graph structure based on the starting and ending points (e.g., tunnel entrances, bridge start and end points) of intersections, interchange entrances / exits, highway ramps, and important road sections within the urban road traffic system. Simultaneously, based on the connectivity between each road segment, road segment connection edges are set between nodes in the weighted directed graph structure, and the static attributes (e.g., number of lanes, design speed, and road grade) and dynamic attributes (e.g., real-time traffic speed, queue length, and accident / construction status) corresponding to each road segment connection edge are obtained. In other words, a dynamic map of the urban road network is constructed using the road segment nodes and connection edges set up in the weighted directed graph structure.

[0025] Furthermore, by using a spatial attention mechanism, the spatial correlation weights between different road segments in the dynamic map of the urban road network are obtained, specifically as follows: ; in: Let the spatial association weights be the edge connecting the i-th road segment and the edge connecting the j-th road segment. Let i be the edge connecting the i-th road segment in the dynamic urban road network graph. Let j be the connecting edge of the j-th road segment in the dynamic urban road network diagram. This is the function for calculating spatial attention.

[0026] Furthermore, based on the road network snapshot (i.e., the state of the road network dynamic map) at each moment in the urban road network dynamic map, the corresponding time-related weights are obtained through a time attention mechanism, specifically: ; in: The time correlation weight between historical time τ and current time t, This is a snapshot of the road network corresponding to a historical moment τ. This is a snapshot of the road network at the current time t. This is a function for calculating temporal attention.

[0027] Furthermore, based on the obtained spatial and temporal correlation weights, these weights are fused layer by layer to obtain the corresponding spatiotemporal joint features, specifically: ; in: It is a spatiotemporal joint feature. Spatial correlation weights, For time-related weights, This is for merging operations (such as splicing, weighted addition, element-wise maximum / average, etc.).

[0028] Specifically, based on the obtained spatiotemporal joint features, these features are combined with the corresponding road segment nodes in the urban road network dynamic map to construct the corresponding road network spatiotemporal feature matrix. Simultaneously, based on the corresponding road segment nodes in the urban road network dynamic map, the rows corresponding to the road network spatiotemporal feature matrix are set, and based on the obtained spatiotemporal joint features, the columns corresponding to the road network spatiotemporal feature matrix are set.

[0029] Step S1.1.2: Vector Generation. This involves extracting relevant historical order data from order logs and aggregating it based on the passenger IDs corresponding to these historical order data. From this, corresponding trip performance data (including actual trip time, estimated trip time, and detour length), service feedback data (including complaint status, complaint type, and complaint handling results), behavioral stability data (including the number of rebookings, cancellations, and temporary destination changes), and travel patterns (including high-frequency departure points, destinations, travel times, and weekday distribution) are extracted. Simultaneously, based on the passenger ID, multiple historical order data corresponding to the same passenger ID are ordered chronologically, and each historical order data is labeled with a corresponding scenario tag (e.g., "peak hours," "rainy days," and "nighttime"). Furthermore, real-time behavioral data is collected from passengers' real-time new orders, including pre-order behavior (including the number of times the pick-up point is modified and the frequency of checking estimated arrival time), recommendation feedback (including the acceptance status of recommended routes), and in-trip behavior (including contacting customer service and changing destinations mid-trip).

[0030] Specifically, based on the acquired historical order data and real-time behavior data, data preprocessing is performed, including missing data handling and normalization. It is worth noting that this embodiment only processes the specific numerical values ​​of the relevant data, and the data preprocessing methods used in this embodiment are all conventional techniques; therefore, they are not specifically described in this embodiment. Simultaneously, based on the preprocessed order data, corresponding time-series features are extracted, including time sensitivity weight, detour tolerance, and dynamic transfer willingness index, specifically: ; in: As a time-sensitivity weight, To allow for detour tolerance, This is the dynamic transfer willingness index. To calculate the total number of orders within the statistical period, This is the indicator value corresponding to a significant delay in the k-th order. , , This is an empirical coefficient. For order indexing, This is the indicator value corresponding to the k-th order being complained about. This is the indication value corresponding to the cancellation of the k-th order. for Let this be the detour rate for a single trip of the k-th order. The value corresponding to the 90th percentile after sorting the detour rate of each order for a single trip within the statistical period in ascending order. For the k-th passenger, this represents the order in which the transfer was actually accepted. For the k-th order whose expected time savings are greater than the threshold, This is an indicator function.

[0031] Furthermore, a variational autoencoder is used to compress and represent the complete behavioral distribution of passengers to extract corresponding implicit preferences, including safety preference, privacy preference, and stability preference. It is worth noting that the implicit preferences in this embodiment are all presented as continuous numerical vectors and are not defined verbally. Simultaneously, the extracted temporal features and implicit preferences are concatenated to construct the corresponding passenger embedding vector. That is, time sensitivity weights, detour tolerance, dynamic transfer willingness index, safety preference, privacy preference, and stability preference are concatenated to construct the corresponding passenger embedding vector.

[0032] Step S1.2: Function Construction. This involves assigning each autonomous vehicle in the fleet to a corresponding intelligent agent, and setting the state space of that agent based on its current location, remaining battery power, passenger information, and current time. The action space of that agent is then set based on each autonomous vehicle's order actions (acceptance / rejection), route selection, and service strategy.

[0033] Furthermore, based on the state space and action space corresponding to each agent, the corresponding system efficiency rewards, detour penalties, expectation violation penalties, and churn risk penalties are determined, specifically as follows: ; in: As a reward for system efficiency, As a penalty for taking a detour, For expected violations of punishment, Penalty for loss of employees This is the empty-running rate weighting coefficient. The full load rate is the weighting factor. Empty running rate For full load rate, Let p be the time sensitivity weight of the p-th passenger. Let p be the actual distance traveled by the p-th passenger. For the shortest path mileage of the p-th passenger, Let p be the actual travel time of the p-th passenger. Let p be the committed travel time for the p-th passenger. Let p be the counterfactual utility difference for the p-th passenger. Let p be the psychological threshold of the p-th passenger. Let be the risk intensity coefficient of the p-th passenger. For indicator functions, Indexing for passengers, This is for passengers currently on the journey to gather.

[0034] It is worth noting that the data involved in the formulas for obtaining system efficiency rewards, detour penalties, expected violation penalties, and churn risk penalties in this embodiment can be specifically set or obtained according to actual needs. For example, the passenger's counterfactual utility difference can be determined based on the difference between the passenger's travel utility in a standalone car dispatch scenario and the passenger's travel utility in a carpooling scenario. Therefore, this embodiment does not provide a specific explanation.

[0035] Furthermore, using the dynamic urban road network map constructed in step S1.1.1, the region label corresponding to each agent is determined, and combined with the scene label corresponding to each agent (such as "peak hours", "rainy day", and "night") to determine the corresponding weight combination for each agent, specifically: ; in: Weighting is assigned to reward system efficiency. To avoid penalty weighting, To prevent the expected violation of the penalty weight, Weighting for penalties based on churn risk. For scene tags, For region labels, This is a mapping function.

[0036] In other words, based on the weight combinations set for different scenarios and regions (which can be specifically set according to actual needs, so they are not specifically described in this embodiment), the corresponding weight combinations are determined through a mapping function. Specifically, this embodiment only uses scenario labels (morning peak, night and holidays) and region labels (business district, low-density area and high-demand area) as examples for illustrative explanation, as shown in Table 1 below (weight illustration table).

[0037] Table 1: Weighting Diagram

[0038] Furthermore, based on the determined weight combination, the corresponding weight magnitude is determined, and this determined weight magnitude is combined with the determined system efficiency reward, detour penalty, expected violation penalty, and churn risk penalty to obtain the corresponding reward function, specifically: ; in: For the reward function, Weighting is assigned to reward system efficiency. To avoid penalty weighting, To prevent the expected violation of the penalty weight, Weighting for penalties based on churn risk. As a reward for system efficiency, As a penalty for taking a detour, For expected violations of punishment, Penalty for the risk of attrition.

[0039] Step S1.3: Function Evaluation. Based on the road network spatiotemporal feature matrix obtained in Step S1.1.1 and the passenger embedding vector obtained in Step S1.1.2, a corresponding set of candidate solutions is constructed (this set can be directly generated by the system, so it is not specifically described in this embodiment). Then, based on the reward function obtained in Step S1.2, the score reward value corresponding to each candidate solution in the set is determined.

[0040] Furthermore, by using a meta-reinforcement learning algorithm, the score reward value corresponding to each candidate solution is combined with the reward function corresponding to each agent (i.e. each autonomous vehicle) in step S1.2, and policy training is performed (which is a conventional algorithm processing method, so it is not specifically described in this embodiment) to directly determine the corresponding optimal candidate solution, including the optimal dispatch strategy and the carpooling path planning scheme.

[0041] In this embodiment, the reward function obtained in step S1.2 is modified using a generative adversarial network model, and the optimal candidate solution determined in step S1.3 is then modified based on the modified reward function. (See reference...) Figure 3 This embodiment provides a dynamic correction method for comfort perception, which specifically includes the following steps: Step S2.1: Loss Calculation. Based on the candidate path set constructed in Step S1.3, determine the road condition parameters (including road surface smoothness, slope, curve curvature, and speed limit), vehicle dynamics parameters (including mass, wheelbase, suspension characteristics, and braking performance), and path geometry information (including steering angle change sequence and speed change sequence) corresponding to each candidate path. Simultaneously, the road condition parameters, vehicle dynamics parameters, and path geometry information corresponding to each candidate path are used as inputs to the generator in the generative adversarial network model, and the corresponding three-dimensional acceleration change sequence is output. Furthermore, based on the passenger's historical physiological feedback data (including the degree of forward lean during sudden braking, heart rate changes during bumps, and subjective comfort ratings), the discriminator in the generative adversarial network model is adversarially trained to output the corresponding comfort loss value (and the range of expected psychological compensation), specifically: ; in: This represents the loss of comfort. Indexing for passengers, Gathering passengers for the current journey, Let p be the space comfort factor for the p-th passenger. , , For three-dimensional acceleration components, , , These are the weighting coefficients for the perceived body position in different directions.

[0042] Step S2.2: Function Compensation. This involves correcting the reward function obtained in Step S1.2 based on the comfort loss value acquired in Step S2.1. Specifically: ; in: This is the initially revised reward function. Assigning a weight to the comfort penalty This represents the loss of comfort. This is the reward function.

[0043] Furthermore, based on the range of expected compensation amounts obtained from the generative adversarial network model output in step S2.1, the corresponding maximum and minimum compensation amounts are determined. These maximum and minimum compensation amounts are then combined to obtain the average compensation amount between the maximum and minimum compensation amounts. This average compensation amount is then used as the corresponding expected compensation amount. Specifically, the initially revised reward function is modified using the determined expected compensation amount to obtain the corresponding final revised reward function, as follows: ; in: For the final revised reward function, This is the initially revised reward function. To compensate for cost weighting, This refers to the amount of compensation expected by the individual.

[0044] Step S2.3: Feedback Correction. Based on the final corrected reward function obtained in Step S2.2, return to Step S1.2 to correct the reward function obtained in Step S1.2. Then, based on the corrected reward function (i.e., the final corrected reward function), execute Step S1.3 to determine the corresponding updated optimal candidate solution. In other words, the optimal candidate solution determined in Step S1.3 is updated using the obtained final corrected reward function, including the updated optimal dispatch strategy and the updated ride-sharing path planning scheme.

[0045] In this embodiment, during shared travel by multiple passengers, the fare discount ratio for each passenger is dynamically adjusted using the Shapley value algorithm and attention mechanism. Furthermore, when the journey time is extended due to carpooling, a corresponding compensation plan is generated based on the comfort loss value and expected compensation range determined in step S2.1. (Reference) Figure 4 This embodiment provides a method for compensating for the gains in a fair game, which specifically includes the following steps: Step S3.1: Contribution Calculation. This involves determining the marginal contribution of each passenger based on the updated optimal dispatch strategy and updated ride-sharing route planning scheme during the ride-sharing process. Simultaneously, based on the detour mileage and time delay for each passenger, the overall contribution of each passenger is determined. This overall contribution is then combined with the dominant position weight of each passenger to determine the discount weight for each passenger. Details are as follows: Step S3.1.1: Contribution Calculation. This involves obtaining the updated optimal dispatch strategy and updated ride-sharing route planning scheme for each passenger during the ride-sharing process. Based on the arrival time corresponding to the updated ride-sharing route planning scheme and the individual travel time for each passenger, the travel time difference for each passenger is determined. Simultaneously, using the Shapley value algorithm, the updated ride-sharing route planning scheme, travel time difference, and the passenger embedding vector constructed in step S1.1.2 are combined to obtain the marginal contribution for each passenger, specifically: ; in: For the p-th passenger in the passenger set The marginal contribution cost below Gathering passengers Total cost when sharing a ride. Let p be the set of passengers. Gather passengers. The new total cost after including the p-th passenger, Index passengers.

[0046] It is worth noting that the total cost in this embodiment is a combination of mileage cost and time cost, for example, based on the passenger set. The total cost of mileage and the total cost of time during ridesharing are combined to determine the passenger set. The total cost when sharing a ride.

[0047] Furthermore, based on the obtained marginal contribution, the overall contribution of each passenger is determined, specifically as follows: ; in: The overall contribution of the p-th passenger. This is the mileage weighting coefficient. For time weighting coefficients, Let p be the time sensitivity weight of the p-th passenger. The extra distance traveled by passengers The time delay caused by the p-th passenger. The time delays caused to passengers Let p be the detour distance caused by the p-th passenger.

[0048] Step S3.1.2: Dominance Identification. This involves concatenating and combining the overall contribution of each passenger, time delay, time sensitivity weight, comfort tolerance threshold, and historical churn probability to construct a co-riding feature vector for each passenger. Simultaneously, an attention mechanism is used to normalize the co-riding feature vector for each passenger to obtain the dominance weight for each passenger. Specifically: ; in: Let p be the weight of the degree of dominance of the p-th passenger. Let be the shared ride feature vector of the p-th passenger. Let v be the ride-sharing feature vector of the v-th passenger. It is an exponential function. , Indexing for passengers, This is for passengers currently on the journey to gather.

[0049] Step S3.1.3: Discount Determination. Based on the overall contribution determined in Step S3.1.1 and the dominance weight obtained in Step S3.1.2, the overall contribution and dominance weight for each passenger are combined to determine the final discount weight for each passenger. Specifically: ; in: Let the final discount weight be for the p-th passenger. The dominant amplification factor (e.g., 0.5-1). Let p be the weight of the degree of dominance of the p-th passenger. The overall contribution of the p-th passenger.

[0050] Step S3.2: Dynamic Compensation. This involves determining which passengers require dynamic compensation by setting time-triggered conditions (i.e., the detour time for each passenger exceeds the corresponding time penalty threshold) and comfort-triggered conditions (i.e., the comfort loss value for each passenger exceeds the corresponding comfort tolerance threshold). It is worth noting that both the time penalty threshold and the comfort tolerance threshold in this embodiment can be specifically set according to actual needs; therefore, they are not specifically described in this embodiment.

[0051] Furthermore, based on the identified passengers requiring dynamic compensation, the expected compensation amount range for each passenger is determined, thus establishing the corresponding maximum and minimum compensation amounts. Simultaneously, based on the passenger embedding vectors constructed in step S1.1.2, passengers are categorized into high-level (e.g., high-frequency travel, high average order value, low complaints, high loyalty), medium-level (e.g., stable travel, occasional carpooling), and low-level (e.g., low-frequency travel, price volatility sensitive, high churn risk). The passenger levels are then combined with the determined expected compensation amount range to determine the corresponding final compensation amount.

[0052] Specifically, when the passenger level is high, the final compensation amount is set based on the maximum compensation amount within the determined range of expected compensation. When the passenger level is low, the final compensation amount is set based on the minimum compensation amount within the determined range of expected compensation. When the passenger level is medium, the maximum and minimum compensation amounts within the expected compensation range are combined with a set compensation bias coefficient to set the corresponding final compensation amount, as follows: ; in: The final compensation amount, To compensate for the bias coefficient, The maximum compensation amount, This is the minimum compensation amount.

[0053] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended embodiments and their equivalents.

Claims

1. A method for optimizing autonomous driving mobility services based on artificial intelligence, characterized in that, Including: S1: Multi-objective planning of data fusion: Through spatiotemporal graph attention network and hybrid feature extraction model, the corresponding spatiotemporal features of road network and multidimensional passenger embedding vector are extracted. At the same time, the global scheduling model is trained through the spatiotemporal features of road network and multidimensional passenger embedding vector. After training, the global scheduling model outputs the corresponding optimal dispatch strategy and carpooling route planning scheme. S2: Dynamic correction of comfort perception: The corresponding comfort loss value is obtained through the generative adversarial network model, and the reward function of the trained global scheduling model is corrected according to the comfort loss value to obtain the corrected optimal dispatch strategy and carpooling path planning scheme. S3: Fair Game Benefit Compensation: When multiple passengers share a ride, the Shapley value algorithm and attention mechanism are used to determine the fare discount ratio for each passenger. When the journey time is extended, the corresponding compensation plan is determined based on the aforementioned comfort loss value and the expected compensation amount.

2. The method for optimizing autonomous driving travel services based on artificial intelligence according to claim 1, characterized in that, The output retrieves the corresponding optimal dispatch strategy and carpooling route planning scheme, including: S1.1: Perception Fusion: Through multiple perception channels set up in parallel, corresponding external environment data and travel user data are collected and acquired. Then, through a spatiotemporal graph attention network and a hybrid feature extraction model, features are extracted from the external environment data and travel user data respectively. S1.2: Function Construction: Each autonomous vehicle in the autonomous vehicle fleet is set as a corresponding intelligent agent. Based on the current location, remaining battery power, passenger information, and current time of the autonomous vehicle, the state space of the intelligent agent is set. Based on the order action, route selection, and service strategy of the autonomous vehicle, the action space of the intelligent agent is set. At the same time, based on the state space and action space, the corresponding reward function is constructed. S1.3: Function Evaluation: Based on the road network spatiotemporal feature matrix and passenger embedding vector, a corresponding candidate solution set is constructed, and the score reward value corresponding to each candidate solution in the candidate solution set is obtained through the reward function. At the same time, the corresponding optimal candidate solution is determined through the meta-reinforcement learning algorithm, including the optimal dispatch strategy and the carpooling route planning scheme.

3. The method for optimizing autonomous driving travel services based on artificial intelligence according to claim 2, characterized in that, Feature extraction was performed on external environment data and travel user data, including: S1.1.1: Feature Extraction: Based on the intersections and road segments of the urban road traffic system, road segment nodes and road segment connection edges with weighted directed graph structures are set to construct the corresponding urban road network dynamic graph. At the same time, through the urban road network dynamic graph, the corresponding spatial association weights and temporal association weights are obtained, and the spatial association weights and temporal association weights are fused to obtain the corresponding spatiotemporal joint features. The spatiotemporal joint features are combined with the road segment nodes to construct the corresponding road network spatiotemporal feature matrix. S1.1.2: Vector Generation: By extracting the corresponding trip performance, service feedback, behavioral stability and travel patterns from the passenger IDs corresponding to historical order data, and by extracting the corresponding real-time behavioral data from real-time new orders, the corresponding temporal features are extracted based on the real-time behavioral data and historical order data. The complete behavioral distribution of passengers is compressed and represented by a variational autoencoder to extract the corresponding implicit preferences. The temporal features and implicit preferences are then concatenated to construct the corresponding passenger embedding vector.

4. The method for optimizing autonomous driving travel services based on artificial intelligence according to claim 2, characterized in that, Based on the state space and action space corresponding to the agent, the system efficiency reward, detour penalty, expected violation penalty, and churn risk penalty corresponding to the agent are determined. At the same time, based on the dynamic map of the urban road network, the region label corresponding to the agent is determined. The region label and scene label are combined to determine the weight combination corresponding to the agent. The weight combination is then combined with the system efficiency reward, detour penalty, expected violation penalty, and churn risk penalty to obtain the corresponding reward function.

5. The method for optimizing autonomous driving travel services based on artificial intelligence according to claim 2, characterized in that, Obtain the revised optimal dispatch strategy and ride-sharing route planning scheme, including: S2.1: Loss Calculation: Based on the road condition parameters, vehicle dynamics parameters and path geometry information corresponding to each candidate path in the candidate solution set, set the input of the generator in the generative adversarial network model and output the corresponding three-dimensional acceleration change sequence. At the same time, based on the passenger's historical physiological feedback data, perform adversarial training on the discriminator in the generative adversarial network model and output the corresponding comfort loss value and psychological expected compensation amount range. S2.2: Function Compensation: Based on the comfort loss value and the expected compensation amount range, the reward function is modified to obtain the corresponding final modified reward function; S2.3: Feedback Correction: Based on the final corrected reward function, feedback step S1.2 is performed to correct the reward function, and based on the corrected reward function, step S1.3 is executed to determine the corresponding updated optimal candidate solution.

6. The method for optimizing autonomous driving travel services based on artificial intelligence according to claim 5, characterized in that, Based on the comfort loss value, the reward function is initially modified to obtain the initially modified reward function, specifically as follows: ； in: This is the initially revised reward function. Weighting for comfort penalty This represents the loss of comfort. For the reward function; Based on the maximum and minimum compensation amounts within the range of expected psychological compensation amounts, the corresponding average compensation amount is determined, and a corresponding expected psychological compensation amount is set. Simultaneously, based on the expected psychological compensation amount, the initially revised reward function is modified to obtain the corresponding final revised reward function, specifically as follows: ； in: For the final revised reward function, This is the initially revised reward function. To compensate for cost weighting, This refers to the amount of compensation expected by the individual.

7. The method for optimizing autonomous driving travel services based on artificial intelligence according to claim 1, characterized in that, The corresponding compensation plan was determined, including: S3.1: Contribution Calculation: Based on the updated optimal dispatch strategy and the updated ride-sharing route planning scheme for each passenger, determine the marginal contribution of each passenger. At the same time, based on the detour mileage and time delay for each passenger, determine the comprehensive contribution of each passenger. Combine the comprehensive contribution with the corresponding dominance weight to determine the discount weight for each passenger. S3.2: Dynamic Compensation: By setting time trigger conditions and comfort trigger conditions, passengers who need dynamic compensation are identified, and based on the range of expected compensation amounts for the corresponding passengers, and based on the passenger embedding vector, passenger levels are divided, and the range of expected compensation amounts and passenger levels are combined to determine the corresponding final compensation amount.

8. The method for optimizing autonomous driving travel services based on artificial intelligence according to claim 7, characterized in that, The discount weight for each passenger is determined, including: S3.1.1: Contribution Calculation: Based on the arrival time and individual travel time corresponding to the updated carpooling route planning scheme, the travel time difference for each passenger is determined. Then, using the Shapley value algorithm, the updated carpooling route planning scheme, the travel time difference, and the passenger embedding vector are combined to obtain the marginal contribution for each passenger. Based on this marginal contribution, the comprehensive contribution of each passenger is determined, specifically as follows: ； in: The overall contribution of the p-th passenger. This is the mileage weighting coefficient. For time weighting coefficients, Let p be the time sensitivity weight of the p-th passenger. The extra distance traveled by passengers The time delay caused by the p-th passenger. The time delays caused to passengers The detour distance caused by the p-th passenger; S3.1.2: Dominance Identification: The overall contribution, time delay, time sensitivity weight, comfort tolerance threshold, and historical churn probability are concatenated and combined to construct a corresponding co-multiplicative feature vector. Then, through an attention mechanism, the dominance weight corresponding to the passenger is obtained, specifically: ； in: Let p be the weight of the degree of dominance of the p-th passenger. Let be the shared ride feature vector of the p-th passenger. Let v be the ride-sharing feature vector of the v-th passenger. It is an exponential function. , Indexing for passengers, Gather passengers for the current journey; S3.1.3: Discount Determination: The final discount weight for the passenger is determined by combining the overall contribution and dominance weights, specifically as follows: ； in: Let the final discount weight be for the p-th passenger. The dominant amplification factor (e.g., 0.5-1). Let p be the weight of the degree of dominance of the p-th passenger. The overall contribution of the p-th passenger.

9. The method for optimizing autonomous driving travel services based on artificial intelligence according to claim 7, characterized in that, The passenger level includes high level, medium level and low level. When the passenger level is high level, the corresponding final compensation amount is set according to the maximum compensation amount in the range of expected compensation amount. When the passenger level is low, the corresponding final compensation amount is set according to the minimum compensation amount in the range of expected compensation amount. When the passenger level is medium, the maximum and minimum compensation amounts within the expected compensation range are combined with a compensation bias coefficient to set the corresponding final compensation amount, specifically: ； in: The final compensation amount, To compensate for the bias coefficient, The maximum compensation amount, This is the minimum compensation amount.

10. An artificial intelligence-based autonomous driving mobility service optimization system, characterized in that, The method for optimizing autonomous driving travel services based on artificial intelligence, as described in any one of claims 1-9, is used.

Images