Intelligent travel assistant system and method based on AIAgent

Abstract

The application discloses an intelligent travel assistant system and method based on an AI Agent, comprising the following steps: obtaining travel-related data and preprocessing to generate current observation data; matching a travel prototype according to the current observation data to generate an action container sequence template; constructing a MuZero model with the current observation data as an input state; performing tree search based on the action container and the sequence template to generate a candidate sequence; constructing a retractable planning chain to realize dynamic adjustment of a path; and outputting a target plan and performing feedback to update a historical prototype library. The application realizes dynamic decision-making and execution linkage of intelligent travel planning, and improves path flexibility and user experience.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an intelligent travel assistant system and method based on AI Agent. Background Technology

[0002] With the development of artificial intelligence, geographic information systems, and mobile internet technologies, travel planning systems have been widely applied in scenarios such as navigation, travel recommendations, and urban commuting. Existing systems typically generate travel plans based on user-inputted origin, destination, or preferences, combined with road network, historical trajectory, or point-of-interest data, using shortest path algorithms, rule-based recommendation models, or static ranking models, which can meet the needs of single-destination or simple route queries.

[0003] In complex travel tasks, existing technologies still have shortcomings. Traditional systems are mostly based on static path calculation, typically generating a complete route all at once before departure, lacking the ability to perceive the continuous state of the travel process and make dynamic decisions. Existing systems lack a phased modeling mechanism for scenarios with multiple points of interest, multiple time constraints, and multiple modes of transportation, making it difficult to uniformly express and coordinately optimize continuous travel links such as walking, public transportation, ride-hailing, railway, and scenic spot reservations, resulting in insufficient integrity and executability of the planning results.

[0004] When traffic congestion occurs, weather changes, service delays occur, or users temporarily change their points of interest, traditional systems typically need to recalculate the global route, lacking mechanisms for quickly replacing local route segments and switching branches. Existing planning results mostly remain at the route display level, with insufficient integration with third-party ticketing, reservation, ride-hailing, and ticket purchase platforms, making it difficult to form a closed loop of planning, execution, and feedback updates. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent travel assistant system and method based on AI Agent. This invention introduces action containerization, journey prototype scheduling, and revocable planning chain structure to achieve dynamic planning and real-time adjustment of travel routes for multiple points of interest, and has the advantages of strong decision-making ability, high route flexibility, and good execution linkage.

[0006] The intelligent travel assistant system and method based on AI Agent according to embodiments of the present invention includes the following steps: Acquire travel-related data, perform preprocessing, generate current observation data, and build an AI Agent; Prototype scheduling involves performing matching calculations in the historical journey prototype library based on current observation data, filtering to obtain a set of candidate journey prototypes, and parsing the set of candidate journey prototypes into corresponding action container sequence templates. Construct the MuZero model, which is a decision model that includes a representation function, a dynamic function, and a prediction function, using the current observation data as the initial input state; Action containerization is performed, and a set of candidate action containers is generated based on the current observation data. Each action container corresponds to a continuous travel sub-path. The set of candidate action containers and the action container sequence template are used together to construct a set of search nodes and input into the MuZero model. Different combinations of action containers are expanded and evaluated through multi-step tree search based on action containers to generate a set of candidate action container sequences. The revocable planning chain is constructed by building a planning chain structure based on the set of candidate action container sequences. Execution conditions, revocation conditions, and alternative action container entries are set for each action container. Several action containers are connected according to the conditions to form a set of candidate planning chains with branch switching relationships. The decision output determines the target planning chain based on the candidate planning chain set and outputs the target planning chain. Feedback updates are performed by updating the journey prototype records in the historical journey prototype library based on the execution results of the target planning chain.

[0007] Optionally, the prototype scheduling specifically includes: Feature extraction is performed on the current observation data to generate a task feature vector; Obtain journey prototype records from the historical journey prototype library, and extract the corresponding prototype feature vector for each journey prototype record; Perform similarity calculations on the task feature vector and each prototype feature vector; select journey prototype records with similarity higher than a preset threshold as a candidate journey prototype set; For each journey prototype record in the candidate journey prototype set, perform structural parsing to extract the stored travel path segment sequence, transportation mode sequence, and corresponding time interval information from the journey prototype record; Perform segmentation processing on each path segment, and convert each path segment into a corresponding action container. The action container contains the path segment, time interval, and combination of transportation modes. Sort and connect each action container according to the path order in the original journey prototype record to generate the corresponding action container sequence template; The action container sequence templates obtained from parsing all candidate journey prototype records are aggregated to form an action container sequence template set.

[0008] Optionally, the MuZero model specifically includes: Establish a decision model that includes a representation function, a dynamic function, and a prediction function. The representation function is used to map the current observation data to the initial hidden state, the dynamic function is used to generate the next hidden state based on the hidden state and the input action, and the prediction function is used to output the policy distribution and value estimation based on the hidden state. The action container is used as the input unit of the dynamic function, and the start location, end location and mode of transportation information contained in the action container are encoded and then used for state transition calculation. During the tree search process, the action container is used as the search extension node. State transitions and value evaluations are performed on different action container sequences, and candidate action container sequences are selected based on the output of the prediction function. The action container sequence template obtained from the prototype scheduling step is used as the initial search branch to insert into the tree search structure. The action container sequence based on the historical journey prototype and the action container sequence generated based on the current observation data jointly participate in the search expansion to form a multi-path decision structure based on action containers.

[0009] Optionally, the action containerization process specifically includes: Based on the current observation data, analyze the user's travel task data, perform matching calculations from the point of interest data, filter out the points of interest, and determine the start and end points of the trip; Arrange the points of interest according to their spatial location to construct a path node sequence that includes the starting point, the points of interest, and the ending point; In the pre-built path network, path search is performed on adjacent nodes in the path node sequence to generate a corresponding continuous location point sequence. The continuous location point sequence is then used to identify the mode of transportation and is divided at the location where the mode of transportation changes to obtain multiple path segments. Each path segment is treated as an action container, and each action container corresponds to the complete travel process from one location to another. Match the action container with the action container sequence template, insert the action container from the action container sequence template into the current action container sequence, and form an expanded action container sequence. The expanded action container sequences are combined according to spatial connectivity, and several action container sequences that satisfy the connectivity constraint between the start and end points are selected. The action container sequence is input into the MuZero model to perform multi-step state transition calculation and value evaluation, and a set of candidate action container sequences is obtained based on the evaluation results.

[0010] Optionally, the construction of the revocable planning chain specifically includes: Set execution conditions for each action container, including the matching relationship between the current time and the corresponding time interval of the action container; Set cancellation conditions for each action container. Cancellation conditions include traffic status changes exceeding a preset threshold, weather status changes exceeding a preset threshold, and user-inputted point of interest change commands. For each action container, set an alternative action container entry. When the cancellation condition is met, select another action container from the candidate action container sequence set whose starting location is the same as the name of the current action container's ending location as the alternative action container. Based on the execution conditions, cancellation conditions, and alternative action container entry points, action containers are connected, and branch connections are established between adjacent action containers in the original action container sequence to form a planning chain structure that includes the main path and alternative paths. During the execution of the planning chain, the execution and cancellation conditions are monitored in real time. When the cancellation condition is met, the system switches to the corresponding alternative action container and continues to execute according to the updated action container sequence to generate a set of candidate planning chains.

[0011] Optionally, the decision output specifically includes: Extract the corresponding action container sequence for each planning chain in the candidate planning chain set, and obtain the cumulative value evaluation result of each action container sequence; The planning chain with the highest cumulative value assessment result is selected as the target planning chain; The action containers in the target planning chain are parsed according to the execution order to extract the corresponding travel routes, modes of transportation, time intervals and location information, and generate structured travel plans. The structured travel plan is converted into visual travel information and output to the user terminal, including route display, time arrangement and switching condition prompts; Based on the transportation mode and location information in the target planning chain, call the external service interface to perform resource retrieval operations and obtain ticketing information, reservation information or service entry information corresponding to each action container; Filter ticketing information, reservation information, or service entry information, select resource links that match the target planning chain, and generate a set of operable links; The set of actionable links is associated with structured travel solutions and displayed to the user's terminal; Upon receiving a user-triggered operation instruction, the system executes the corresponding ticket purchase or reservation operation based on the set of operable links, and updates the execution status of the target planning chain.

[0012] Optionally, the feedback update specifically includes: During the execution of the target planning chain, execution result data is collected, including the actual execution path, actual time interval, execution status of transportation mode, and user operation behavior of each action container. The execution result data is compared with the original action container sequence in the target planning chain to calculate the path deviation and time deviation; When both the path deviation and the time deviation are less than the preset threshold, the corresponding action container sequence is updated and written into the historical journey prototype library as a new journey prototype record. When the path deviation or time deviation exceeds the preset threshold, the journey prototype record corresponding to the target planning chain in the historical journey prototype library is updated, and the execution result data replaces the path and time information in the original journey prototype record. Perform validity checks on each journey prototype record in the historical journey prototype library. If a journey prototype record is called less than a preset threshold number of times within a preset time window or the deviation of the corresponding execution result data exceeds a preset number of times consecutively, the journey prototype record will be deleted from the historical journey prototype library.

[0013] Optionally, the AI ​​Agent includes: The core agent performs semantic parsing on user travel task data in travel-related data and incorporates the parsing results into the current observation data; The planning agent executes prototype scheduling and action containerization processing. Based on the current observation data, it selects a set of candidate journey prototypes from the historical journey prototype library, generates action container sequence templates, and generates a set of candidate action containers based on the current observation data. The planning agent constructs a search node set by combining the candidate action container set and the action container sequence template, and inputs it into the MuZero model to generate a candidate action container sequence set through multi-step tree search based on action containers; The verification agent performs time interval matching, spatial connection relationship verification, and action container execution condition verification on the candidate action container sequence set or candidate planning chain set to determine the target planning chain; During the execution of the target planning chain, the accompanying agent detects the execution and cancellation conditions of each action container. When the cancellation condition is met, it triggers the corresponding alternative action container entry and updates the subsequent execution status of the target planning chain.

[0014] An AI Agent-based intelligent travel assistant system according to an embodiment of the present invention includes: The user interaction module is used to receive travel requests submitted by users and display itinerary plans and execution status to users; The data acquisition module is used to provide data services, including points of interest, traffic, weather, ticketing, and user behavior, to the AI ​​Agent and decision planning module, and to generate current observation data. The prototype scheduling module is used to call the personalized recommendation model to obtain points of interest based on the current observation data and to match action container sequence templates from the historical journey prototype library. The decision planning module is used to build the MuZero model, perform multi-step state transition calculations and tree search on the action container set and action container sequence template, generate a candidate action container sequence set and build a revocable planning chain; The third-party service call module is used to call external service interfaces based on the transportation mode information and location information in the target planning chain, obtain ticketing information, reservation information or service entry information, and generate operable links. The execution and feedback module is used to execute ticket purchase or reservation operations based on user-triggered actions, update the execution status of the target planning chain, update the journey prototype records in the historical journey prototype library based on the execution results, and feed the execution results back to the user interaction module.

[0015] The beneficial effects of this invention are: (1) This invention divides the continuous travel process of multiple points of interest and multiple modes of transportation into several structured travel sub-paths by containerizing the action, so that complex paths can participate in the calculation in a unified form, reducing invalid path combinations, reducing search space and computational complexity, and improving planning efficiency.

[0016] (2) This invention uses a journey prototype scheduling mechanism and MuZero decision model to convert historical high-quality travel experience into action container sequence templates, and combines current observation data to perform multi-step state transition calculations and tree search, thereby improving the quality of initial planning, path stability and the overall rationality of multi-stage travel decisions.

[0017] (3) The present invention uses a revocable planning chain structure to set execution conditions, cancellation conditions and alternative action container entry in each action container, so that the system can switch local paths when traffic, weather or user plans change, reduce overall replanning calls, reduce the load of repeated access to third-party service interfaces, and improve the user interaction experience in dynamic travel scenarios. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the AI ​​Agent-based intelligent travel assistant method proposed in this invention; Figure 2 This is a structural diagram of the AI ​​Agent-based intelligent travel assistant system proposed in this invention; Figure 3 This is a flowchart illustrating the Agent execution process of the AI ​​Agent-based intelligent travel assistant method proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-3 The intelligent travel assistant system and method based on AI Agent includes the following steps: Acquire travel-related data, perform preprocessing, and generate current observation data; Prototype scheduling involves performing matching calculations in the historical journey prototype library based on current observation data, filtering to obtain a set of candidate journey prototypes, and parsing the set of candidate journey prototypes into corresponding action container sequence templates. Construct the MuZero model, which is a decision model that includes a representation function, a dynamic function, and a prediction function, using the current observation data as the initial input state; Action containerization is performed, and a set of candidate action containers is generated based on the current observation data. Each action container corresponds to a continuous travel sub-path. The set of candidate action containers and the action container sequence template are used together to construct a set of search nodes and input into the MuZero model. Different combinations of action containers are expanded and evaluated through multi-step tree search based on action containers to generate a set of candidate action container sequences. The revocable planning chain is constructed by building a planning chain structure based on the set of candidate action container sequences. Execution conditions, revocation conditions, and alternative action container entries are set for each action container. Several action containers are connected according to the conditions to form a set of candidate planning chains with branch switching relationships. The decision output determines the target planning chain based on the candidate planning chain set and outputs the target planning chain. Feedback updates are performed by updating the journey prototype records in the historical journey prototype library based on the execution results of the target planning chain.

[0021] In this embodiment, travel-related data is acquired, including user travel task data, terminal location data, and environmental status data. User travel task data includes departure point, destination, and time constraint information. Terminal location data is used to characterize the user's current spatial location. Environmental status data includes traffic status data and weather status data. Preprocessing operations are performed on the travel-related data, including data cleaning, outlier removal, time alignment, and coordinate transformation. Data from different sources are uniformly converted into a structured data format to generate current observation data.

[0022] In this embodiment, prototype scheduling specifically includes: Feature extraction is performed on the current observation data to generate a task feature vector; Obtain journey prototype records from the historical journey prototype library, extract the corresponding prototype feature vector for each journey prototype record, and use the same feature dimension as the task feature vector. Perform similarity calculation on the task feature vector and each prototype feature vector; select the journey prototype records with similarity higher than the preset threshold as the candidate journey prototype set; the similarity calculation adopts the cosine similarity calculation method; the preset threshold is set to 0.75; if the number of candidate journey prototypes exceeds 5, select the top 5 journey prototype records in descending order of similarity. For each journey prototype record in the candidate journey prototype set, perform structural parsing to extract the stored travel path segment sequence, transportation mode sequence, and corresponding time interval information from the journey prototype record; Each path segment is segmented, and each path segment is converted into a corresponding action container. The action container contains the path segment, time interval, and combination of transportation modes. The segmentation is divided according to the change of transportation modes. When the transportation mode changes, a new path segment is forcibly divided. The action containers are sorted and connected according to the path order in the original journey prototype record to generate the corresponding action container sequence template. Each action container includes a start point, an end point, a mode of transportation identifier and a corresponding path set. The start point is the starting point of the path segment in the action container, and the end point is the ending point of the path segment in the action container. The path set includes multiple consecutive spatial location points to completely describe the travel process from the start point to the end point. The action container sequence templates obtained from parsing all candidate journey prototype records are aggregated to form an action container sequence template set.

[0023] In this embodiment, the MuZero model specifically includes: Establish a decision model that includes a representation function, a dynamic function, and a prediction function. The representation function is used to map the current observation data to the initial hidden state, the dynamic function is used to generate the next hidden state based on the hidden state and the input action, and the prediction function is used to output the policy distribution and value estimation based on the hidden state. The action container is used as the input unit of the dynamic function, and the start location, end location and mode of transportation information contained in the action container are encoded and then used for state transition calculation. During the tree search process, the action container is used as the search extension node. State transitions and value evaluations are performed on different action container sequences, and candidate action container sequences are selected based on the output of the prediction function. The action container sequence template obtained from the prototype scheduling step is used as the initial search branch to insert into the tree search structure. The action container sequence based on the historical journey prototype and the action container sequence generated based on the current observation data jointly participate in the search expansion to form a multi-path decision structure based on action containers.

[0024] In this embodiment, the action containerization process specifically includes: Based on the current observation data, analyze the user's travel task data, perform matching calculations from the point of interest (POI) data, filter out POIs, and determine the travel start and end points; based on the current observation data, analyze the user's travel task data, call the preset personalized recommendation model to perform matching calculations on the POI data; the travel start point is determined by the location data, and the end point is the POI corresponding to the preset rest location. Arrange the points of interest according to their spatial location to construct a path node sequence containing the starting point, the points of interest, and the ending point; sort the points of interest from near to far according to their spatial distance, using the shortest path distance as the sorting criterion, and generate a single-order path node sequence. In a pre-constructed path network, path search is performed on adjacent nodes in the path node sequence to generate a corresponding continuous location point sequence. The continuous location point sequence is then used to identify the mode of transportation. The sequence is divided at locations where the mode of transportation changes to obtain multiple path segments. The path search uses the K-shortest path algorithm to generate no more than 5 paths, and the continuous location point sequence corresponding to the shortest path is selected. The mode of transportation identification is determined based on the traffic type labels in the path network. Each path segment is treated as an action container, and each action container corresponds to a complete travel process from one location to another, including the starting location, the ending location, and the mode of transportation. The process involves matching action containers with action container sequence templates, inserting action containers from the action container sequence templates into the current action container sequence to form an expanded action container sequence; reading each template action container from the action container sequence template one by one, and extracting the start and end points of the template action containers; performing a matching judgment on each action container in the current action container sequence, matching the end point of the current action container with the start point of the template action container by name; inserting the template action container after the corresponding action container in the current action container sequence at the position where the names match; performing a sequence check on the inserted action container sequence, checking whether the interest point at the end point of the inserted action container is consistent with the interest point name of the path not planned based on the spatial connection relationship, removing inconsistent insertion results, and generating an expanded action container sequence. The expanded action container sequence is combined according to the spatial connection relationship, and several action container sequences that satisfy the connection constraint between the start and end points are selected. The connection constraint between the start and end points means that the end point of the previous action container is the same as the start point of the next action container. The action container sequence is input into the MuZero model to perform multi-step state transition calculation and value evaluation, and a set of candidate action container sequences is obtained based on the evaluation results.

[0025] In this embodiment, the construction of the revocable planning chain specifically includes: Set execution conditions for each action container. The execution conditions include the matching relationship between the current time and the corresponding time interval of the action container. The matching relationship is that the current time is between the start time and the end time of the corresponding time interval of the action container. For each action container, a cancellation condition is set. The cancellation condition includes traffic condition changes exceeding a preset threshold, weather condition changes exceeding a preset threshold, and user-inputted point of interest change command. Traffic condition changes exceeding the preset threshold are when the expected increase in road segment travel time exceeds 20%, and weather condition changes exceeding the preset threshold are when the rainfall level or wind force level increases to a preset level or above. For each action container, set an alternative action container entry. When the cancellation condition is met, select another action container from the candidate action container sequence set whose starting location is the same as the name of the current action container's ending location as the alternative action container. Based on the execution conditions, cancellation conditions, and alternative action container entry points, action containers are connected, and branch connections are established between adjacent action containers in the original action container sequence to form a planning chain structure that includes the main path and alternative paths. During the execution of the planning chain, the execution and cancellation conditions are monitored in real time. When the cancellation condition is met, the system switches to the corresponding alternative action container and continues to execute according to the updated action container sequence to generate a set of candidate planning chains.

[0026] By setting execution conditions, cancellation conditions, and alternative action container entry points in each action container, rapid switching of local paths can be achieved, avoiding recalculation of the entire path, thereby reducing system response time and computational resource consumption.

[0027] In this embodiment, the decision output specifically includes: Extract the corresponding action container sequence for each planning chain in the candidate planning chain set, and obtain the cumulative value assessment result of each action container sequence; the cumulative value assessment result is the sum of the value assessment results output by each action container in the MuZero model; The planning chain with the highest cumulative value assessment result is selected as the target planning chain; The action containers in the target planning chain are parsed according to the execution order to extract the corresponding travel routes, modes of transportation, time intervals and location information, and generate structured travel plans. The structured travel plan is converted into visual travel information and output to the user terminal, including route display, time arrangement and switching condition prompts; Based on the transportation mode and location information in the target planning chain, external service interfaces are called to perform resource retrieval operations to obtain ticketing information, reservation information, or service entry information corresponding to each action container; the external service interfaces are third-party ticketing interfaces or reservation interfaces, and the resource retrieval operation adopts a query method based on keyword matching; Filter ticketing information, reservation information, or service entry information, select resource links that match the target planning chain, and generate a set of operable links; The set of actionable links is associated with the structured travel solution and displayed on the user's terminal; the association is displayed by showing clickable links at the corresponding action container position; When a user triggers an operation instruction, the corresponding ticket purchase or reservation operation is executed based on the set of operable links, and the execution status of the target planning chain is updated. The user trigger operation instruction is a click operation or a confirmation operation, and the execution status includes non-executed and completed status.

[0028] In this embodiment, the feedback update specifically includes: During the execution of the target planning chain, execution result data is collected, including the actual execution path, actual time interval, execution status of transportation mode, and user operation behavior of each action container. The execution result data is compared with the original action container sequence in the target planning chain to calculate the path deviation and time deviation; When both path deviation and time deviation are less than the preset threshold, the corresponding action container sequence is updated and written to the historical journey prototype library as a new journey prototype record; path deviation is the number of times path replanning is triggered during execution; time deviation is the time difference between the actual arrival time and the planned arrival time. When the path deviation or time deviation exceeds the preset threshold, the journey prototype record corresponding to the target planning chain in the historical journey prototype library is updated, and the execution result data replaces the path and time information in the original journey prototype record. The update process is to overwrite and update the path and time information of the corresponding action container in the original journey prototype record. The path deviation threshold is 2 times, and the time deviation threshold is 1 hour. Perform validity checks on each journey prototype record in the historical journey prototype library. If a journey prototype record is called less than a preset number of times within a preset time window or the deviation of the corresponding execution result data exceeds a preset number of times consecutively, the journey prototype record will be deleted from the historical journey prototype library. The preset time window is 30 days, the preset number of times threshold is 3 times, and the deviation exceeds the preset number of times consecutively is when the path deviation or time deviation in the execution result exceeds the corresponding threshold for 3 consecutive times.

[0029] In this embodiment, the AI ​​Agent includes: The core agent performs semantic parsing on user travel task data in travel-related data and incorporates the parsing results into the current observation data; The planning agent executes prototype scheduling and action containerization processing. Based on the current observation data, it selects a set of candidate journey prototypes from the historical journey prototype library, generates action container sequence templates, and generates a set of candidate action containers based on the current observation data. The planning agent constructs a search node set by combining the candidate action container set and the action container sequence template, and inputs it into the MuZero model to generate a candidate action container sequence set through multi-step tree search based on action containers; The verification agent performs time interval matching, spatial connection relationship verification, and action container execution condition verification on the candidate action container sequence set or candidate planning chain set to determine the target planning chain; During the execution of the target planning chain, the accompanying agent detects the execution and cancellation conditions of each action container. When the cancellation condition is met, it triggers the corresponding alternative action container entry and updates the subsequent execution status of the target planning chain.

[0030] In the operation of the system of this invention, the AI ​​Agent does not work independently as a single functional unit, but is embedded among modules such as user interaction, data collection, route planning, execution linkage, and feedback update, undertaking the roles of task understanding, decision coordination, and process control. During the user input phase, the AI ​​Agent converts natural language travel needs, interests, preferences, and temporary change instructions into computable task constraints. During the data processing phase, the AI ​​Agent semantically integrates data on points of interest, traffic, weather, ticketing, and user behavior to form the current observation state for planning decisions. During the route generation phase, the AI ​​Agent calls the historical journey prototype library and the decision planning module to uniformly map the historical route structure and current route segments into action container sequences, coordinating the MuZero model to complete candidate solution search. During the solution execution phase, the AI ​​Agent determines whether to continue execution, partially replace, or trigger third-party service calls based on the execution and cancellation conditions in the planning chain. During the feedback update phase, the AI ​​Agent writes back the actual execution deviation and user operation results to the historical journey prototype library. Therefore, the responsibilities of each Agent are not simply corresponding to fixed modules, but rather involve scheduling and connecting multiple modules at different business stages, enabling the system to form a continuous closed loop from requirement analysis to dynamic execution and then to experience updating.

[0031] An AI Agent-based intelligent travel assistant system according to an embodiment of the present invention includes: The user interaction module is used to receive travel requests submitted by users and display itinerary plans and execution status to users; The data acquisition module is used to provide data services, including points of interest, traffic, weather, ticketing, and user behavior, to the AI ​​Agent and decision planning module, and to generate current observation data. The prototype scheduling module is used to call the personalized recommendation model to obtain points of interest based on the current observation data and to match action container sequence templates from the historical journey prototype library. The decision planning module is used to build the MuZero model, perform multi-step state transition calculations and tree search on the action container set and action container sequence template, generate a candidate action container sequence set and build a revocable planning chain; The third-party service call module is used to call external service interfaces based on the transportation mode information and location information in the target planning chain, obtain ticketing information, reservation information or service entry information, and generate operable links. The execution and feedback module is used to execute ticket purchase or reservation operations based on user-triggered actions, update the execution status of the target planning chain, update the journey prototype records in the historical journey prototype library based on the execution results, and feed the execution results back to the user interaction module.

[0032] In this invention, the intelligent travel assistant system does not simply configure the various functional modules in parallel, but rather forms a collaborative operating architecture around the generation, decision-making, execution, and feedback processes of complex travel tasks. The user interaction module serves as the input / output entry and exit point for travel needs and execution status, transmitting user-submitted information such as departure point, destination, interests, time constraints, and temporary adjustments to the system. The data acquisition module continuously gathers data on points of interest, traffic, weather, ticketing, reservation resources, and user behavior, generating current observation data that can be used for subsequent calculations. Based on the current observation data, the prototype scheduling module calls upon the historical journey prototype library to convert similar travel experiences into action container sequence templates, providing experience guidance for subsequent path combinations. The decision planning module abstracts the current path segment into an action container, inputs the action container sequence template into the MuZero decision model, generates candidate action container sequences through multi-step tree search, and constructs a revocable planning chain based on execution conditions, cancellation conditions, and alternative entry points. The third-party service invocation module obtains ticketing, reservation, or service entry points based on the location and transportation information in the target planning chain, enabling the planning results to enter the actual execution stage. The execution and feedback module updates the state of the target planning chain based on user operations and actual travel results, and reverse-corrects the historical journey prototype library. Thus, this invention achieves continuous collaboration between action containers, journey prototype scheduling, the MuZero decision model, and revocable planning chains in a complete system environment.

[0033] Example 1: To verify the feasibility of this invention in practice, it was applied to a multi-point-of-interest (POI) continuous travel planning scenario. In this scenario, users need to visit multiple POIs within a limited time, combining real-time traffic conditions, weather changes, and temporary adjustment needs to complete the overall itinerary arrangement. In existing technologies, traditional shortest path methods only calculate paths for a single destination and cannot handle the problem of multiple POI combinations; while recommendation system-based methods can provide POI filtering, they lack path structure optimization capabilities; reinforcement learning path planning methods, although possessing certain decision-making capabilities, lack flexible adjustment mechanisms during actual execution and are difficult to cope with dynamic changes.

[0034] In this embodiment, user-input travel task data is acquired, and current observation data is generated by combining location data and environmental data. Multiple points of interest are filtered using a pre-set personalized recommendation model, constructing a path node sequence including the starting point, intermediate points of interest, and ending point. Path search is performed within the path network to generate a continuous sequence of location points, and path segments are divided based on changes in transportation mode. Each path segment is constructed as an action container. An action container sequence template is obtained by combining it with a historical journey prototype library, and the template is inserted into the current path structure to enhance the path structure.

[0035] During the decision-making phase, a MuZero model is constructed, using action containers as the basic decision-making unit to perform multi-step state transition calculations and tree search, generating multiple candidate action container sequences. By introducing a reversible planning chain structure, execution and reversal conditions are set for each action container, and alternative action container entry points are configured. This allows the system to adjust only local paths when faced with traffic congestion, weather changes, or temporary changes in user points of interest, without requiring overall replanning. Finally, a target planning chain is generated, outputting a structured travel plan. External service interfaces are called to obtain ticketing and reservation information, enabling automated execution. During execution, execution result data is collected in real time, and the historical journey prototype library is updated and deleted, forming a continuous optimization mechanism.

[0036] To verify the beneficial effects of this invention, its method was compared with traditional shortest path methods, recommendation system-based methods, and reinforcement learning path planning methods. Multiple experiments were conducted under the same travel task conditions, and indicators such as average planning time, number of path adjustments, user satisfaction, execution success rate, and time deviation were statistically analyzed. The results show that this invention demonstrates advantages in all indicators.

[0037] Table 1: Comparison of Path Planning Methods

[0038] As shown in Table 1, the method of this invention outperforms existing technologies in overall performance. Regarding planning efficiency, the average planning time is reduced from 4.8 seconds using traditional methods to 2.4 seconds, a reduction of approximately 50%, indicating that the combination of action containers and the MuZero model effectively reduces search complexity. In terms of path stability, the number of path adjustments is reduced from 3 to 1, indicating improved initial planning quality. User satisfaction increases from 82% to 92%, demonstrating that the generated travel plans better meet user needs. Regarding execution success rate, this invention achieves 95%, higher than other methods, indicating higher reliability in actual execution.

[0039] In summary, this invention, through structured decision-making and dynamic adjustment mechanisms, realizes the transformation of travel planning from static calculation to intelligent decision-making and execution linkage, thereby improving the overall performance of the system.

[0040] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An AI Agent based intelligent travel assistant method, characterized in that, Includes the following steps: Acquire travel-related data, perform preprocessing, generate current observation data, and build an AI Agent; Prototype scheduling involves performing matching calculations in the historical journey prototype library based on current observation data, filtering to obtain a set of candidate journey prototypes, and parsing the set of candidate journey prototypes into corresponding action container sequence templates. Construct the MuZero model, which is a decision model that includes a representation function, a dynamic function, and a prediction function, using the current observation data as the initial input state; Action containerization is performed, and a set of candidate action containers is generated based on the current observation data. Each action container corresponds to a continuous travel sub-path. The set of candidate action containers and the action container sequence template are used together to construct a set of search nodes and input into the MuZero model. Different combinations of action containers are expanded and evaluated through multi-step tree search based on action containers to generate a set of candidate action container sequences. The revocable planning chain is constructed by building a planning chain structure based on the set of candidate action container sequences. Execution conditions, revocation conditions, and alternative action container entries are set for each action container. Several action containers are connected according to the conditions to form a set of candidate planning chains with branch switching relationships. The decision output determines the target planning chain based on the candidate planning chain set and outputs the target planning chain. Feedback updates are performed by updating the journey prototype records in the historical journey prototype library based on the execution results of the target planning chain. 2.The AI Agent based intelligent travel assistant method according to claim 1, wherein, The prototype scheduling specifically includes: Feature extraction is performed on the current observation data to generate a task feature vector; Obtain journey prototype records from the historical journey prototype library, and extract the corresponding prototype feature vector for each journey prototype record; Perform similarity calculations on the task feature vector and each prototype feature vector; select journey prototype records with similarity higher than a preset threshold as a candidate journey prototype set; For each journey prototype record in the candidate journey prototype set, perform structural parsing to extract the stored travel path segment sequence, transportation mode sequence, and corresponding time interval information from the journey prototype record; Perform segmentation processing on each path segment, and convert each path segment into a corresponding action container. The action container contains the path segment, time interval, and combination of transportation modes. Sort and connect each action container according to the path order in the original journey prototype record to generate the corresponding action container sequence template; The action container sequence templates obtained from parsing all candidate journey prototype records are aggregated to form an action container sequence template set. 3.The AI Agent based intelligent travel assistant method according to claim 2, wherein, The MuZero model specifically includes: Establish a decision model that includes a representation function, a dynamic function, and a prediction function. The representation function is used to map the current observation data to the initial hidden state, the dynamic function is used to generate the next hidden state based on the hidden state and the input action, and the prediction function is used to output the policy distribution and value estimation based on the hidden state. The action container is used as the input unit of the dynamic function, and the start location, end location and mode of transportation information contained in the action container are encoded and then used for state transition calculation. During the tree search process, the action container is used as the search extension node. State transitions and value evaluations are performed on different action container sequences, and candidate action container sequences are selected based on the output of the prediction function. The action container sequence template obtained from the prototype scheduling step is used as the initial search branch to insert into the tree search structure. The action container sequence based on the historical journey prototype and the action container sequence generated based on the current observation data jointly participate in the search expansion to form a multi-path decision structure based on action containers. 4.The AI Agent based intelligent travel assistant method according to claim 3, wherein, The action containerization process specifically includes: Based on the current observation data, analyze the user's travel task data, perform matching calculations from the point of interest data, filter out the points of interest, and determine the start and end points of the trip; Arrange the points of interest according to their spatial location to construct a path node sequence that includes the starting point, the points of interest, and the ending point; In the pre-built path network, path search is performed on adjacent nodes in the path node sequence to generate a corresponding continuous location point sequence. The continuous location point sequence is then used to identify the mode of transportation and is divided at the location where the mode of transportation changes to obtain multiple path segments. Each path segment is treated as an action container, and each action container corresponds to the complete travel process from one location to another. Match the action container with the action container sequence template, insert the action container from the action container sequence template into the current action container sequence, and form an expanded action container sequence. The expanded action container sequences are combined according to spatial connectivity, and several action container sequences that satisfy the connectivity constraint between the start and end points are selected. The action container sequence is input into the MuZero model to perform multi-step state transition calculation and value evaluation, and a set of candidate action container sequences is obtained based on the evaluation results. 5.The AI Agent based intelligent travel assistant method according to claim 4, wherein, The construction of the revocable planning chain specifically includes: Set execution conditions for each action container, including the matching relationship between the current time and the corresponding time interval of the action container; Set cancellation conditions for each action container. Cancellation conditions include traffic status changes exceeding a preset threshold, weather status changes exceeding a preset threshold, and user-inputted point of interest change commands. For each action container, set an alternative action container entry. When the cancellation condition is met, select another action container from the candidate action container sequence set whose starting location is the same as the name of the current action container's ending location as the alternative action container. Based on the execution conditions, cancellation conditions, and alternative action container entry points, action containers are connected, and branch connections are established between adjacent action containers in the original action container sequence to form a planning chain structure that includes the main path and alternative paths. During the execution of the planning chain, the execution and cancellation conditions are monitored in real time. When the cancellation condition is met, the system switches to the corresponding alternative action container and continues to execute according to the updated action container sequence to generate a set of candidate planning chains. 6.The AI Agent based intelligent travel assistant method according to claim 5, wherein, The decision output specifically includes: Extract the corresponding action container sequence for each planning chain in the candidate planning chain set, and obtain the cumulative value evaluation result of each action container sequence; The planning chain with the highest cumulative value assessment result is selected as the target planning chain; The action containers in the target planning chain are parsed according to the execution order to extract the corresponding travel routes, modes of transportation, time intervals and location information, and generate structured travel plans. The structured travel plan is converted into visual travel information and output to the user terminal, including route display, time arrangement and switching condition prompts; Based on the transportation mode and location information in the target planning chain, call the external service interface to perform resource retrieval operations and obtain ticketing information, reservation information or service entry information corresponding to each action container; Filter ticketing information, reservation information, or service entry information, select resource links that match the target planning chain, and generate a set of operable links; The set of actionable links is associated with structured travel solutions and displayed to the user's terminal; Upon receiving a user-triggered operation instruction, the system executes the corresponding ticket purchase or reservation operation based on the set of operable links, and updates the execution status of the target planning chain. 7.The AI Agent based intelligent travel assistant method according to claim 6, wherein, The feedback update specifically includes: During the execution of the target planning chain, execution result data is collected, including the actual execution path, actual time interval, execution status of transportation mode, and user operation behavior of each action container. The execution result data is compared with the original action container sequence in the target planning chain to calculate the path deviation and time deviation; When both the path deviation and the time deviation are less than the preset threshold, the corresponding action container sequence is updated and written into the historical journey prototype library as a new journey prototype record. When the path deviation or time deviation exceeds the preset threshold, the journey prototype record corresponding to the target planning chain in the historical journey prototype library is updated, and the execution result data replaces the path and time information in the original journey prototype record. Perform validity checks on each journey prototype record in the historical journey prototype library. If a journey prototype record is called less than a preset threshold number of times within a preset time window or the deviation of the corresponding execution result data exceeds a preset number of times consecutively, the journey prototype record will be deleted from the historical journey prototype library. 8.The AI Agent based intelligent travel assistant method according to claim 7, wherein, The AIAgent includes: The core agent performs semantic parsing on user travel task data in travel-related data and incorporates the parsing results into the current observation data; The planning agent executes prototype scheduling and action containerization processing. Based on the current observation data, it selects a set of candidate journey prototypes from the historical journey prototype library, generates action container sequence templates, and generates a set of candidate action containers based on the current observation data. The planning agent constructs a search node set by combining the candidate action container set and the action container sequence template, and inputs it into the MuZero model to generate a candidate action container sequence set through multi-step tree search based on action containers; The verification agent performs time interval matching, spatial connection relationship verification, and action container execution condition verification on the candidate action container sequence set or candidate planning chain set to determine the target planning chain; During the execution of the target planning chain, the accompanying agent detects the execution and cancellation conditions of each action container. When the cancellation condition is met, it triggers the corresponding alternative action container entry and updates the subsequent execution status of the target planning chain.

9. An AI Agent-based intelligent travel assistant system, applied to the AI ​​Agent-based intelligent travel assistant method according to any one of claims 1 to 8, characterized in that, include: The user interaction module is used to receive travel requests submitted by users and display itinerary plans and execution status to users; The data acquisition module is used to provide data services, including points of interest, traffic, weather, ticketing, and user behavior, to the AI ​​Agent and decision planning module, and to generate current observation data. The prototype scheduling module is used to call the personalized recommendation model to obtain points of interest based on the current observation data and to match action container sequence templates from the historical journey prototype library. The decision planning module is used to build the MuZero model, perform multi-step state transition calculations and tree search on the action container set and action container sequence template, generate a candidate action container sequence set and build a revocable planning chain; The third-party service call module is used to call external service interfaces based on the transportation mode information and location information in the target planning chain, obtain ticketing information, reservation information or service entry information, and generate operable links. The execution and feedback module is used to execute ticket purchase or reservation operations based on user-triggered actions, update the execution status of the target planning chain, update the journey prototype records in the historical journey prototype library based on the execution results, and feed the execution results back to the user interaction module.

Images