Urban intelligent travel path recommendation method and device, storage medium and computer equipment

Abstract

The invention provides a city intelligent travel path recommendation method and device, a storage medium and computer equipment. The method comprises the following steps: analyzing a travel demand instruction of a user based on a large language model to obtain a structured current travel element; performing multi-modal planning processing according to the current travel element to obtain a plurality of candidate paths; and sorting and screening the plurality of candidate paths according to historical travel information of the user to obtain a recommended path. The recommended path obtained by the method not only conforms to the current travel demand, but also conforms to the personalized preference, and the method has relatively strong interaction capability and reasoning capability.

Description

Technical Field

[0001] This application belongs to the field of information technology, and specifically relates to a method, device, computer-readable storage medium, and computer equipment for recommending intelligent urban travel routes. Background Technology

[0002] The evolution of urban mobility planning systems has progressed from shortest path calculation based on static map data to intelligent recommendations incorporating real-time information and user preferences. Early systems, such as navigation tools based on Dijkstra's algorithm, laid the theoretical foundation for path search. With technological advancements, researchers have introduced more efficient algorithms (such as shrinking hierarchies) to handle large-scale road networks and utilized real-time traffic data to achieve dynamic replanning, significantly improving the accuracy and usability of paths.

[0003] To enhance the intelligence level of the system, research has further shifted its focus to personalized and multimodal planning. On the one hand, researchers have developed models based on collaborative filtering, matrix factorization, and deep learning (such as recurrent neural networks (RNNs) and graph neural networks (GNNs)) to learn users' travel preferences from their historical trajectory data, thereby achieving personalized route ranking and recommendation. On the other hand, for complex urban transportation networks, multimodal route planning research aims to address the challenge of seamlessly connecting various modes of transportation such as walking, public transportation, subways, and cycling to generate integrated travel solutions.

[0004] Existing systems (including academic prototypes and industrial applications such as Google Maps) still have common limitations in terms of interactive intelligence:

[0005] (1) Rigid interaction: It relies on structured form input and has difficulty understanding the complex and multi-constrained requirements of natural language expression.

[0006] (2) Insufficient reasoning ability: lack of common sense and contextual reasoning, unable to proactively suggest that users avoid peak hours or understand implicit intentions.

[0007] (3) Limited personalization: It relies heavily on explicit historical data, making it difficult to capture implicit preferences and unable to continuously learn and evolve in dialogue.

[0008] (4) Service fragmentation: Path planning, POI retrieval, real-time information query and other functional modules are independent and lack a unified intelligent agent for autonomous scheduling and task integration.

[0009] In recent years, large language models have demonstrated emergent intelligence, exhibiting unprecedented capabilities in natural language understanding and generation, knowledge reasoning, and code generation. Researchers have begun to explore the potential of large language models (LLMs) as autonomous agents, enhancing them through mechanisms such as the introduction of memory modules, planning capabilities, and tool usage. For example, Wang et al. and Liu et al. equipped LLMs with external memory modules, thereby endowing the models with the potential for continuous growth. In terms of planning and reasoning, the Chain-of-Track (CoT) method and the ReAct framework improve the model's planning ability through stepwise reasoning; Tree-of-Track (ToT) and Graph-of-Track (GoT) further introduce multi-path reasoning mechanisms to ensure the consistency and correctness of the reasoning process. In addition, Self-Refine and Reflexion methods guide LLMs to reflect on their own errors, thereby improving the success rate of solving subsequent problems.

[0010] In the specific field of urban route planning, as mentioned earlier, there exists a mature and powerful "toolset"—various efficient route planning algorithms. However, these powerful algorithms are currently operating independently, lacking an "operator" capable of understanding and organically integrating the user's high-level intentions. Therefore, combining the general interaction and task planning capabilities of large language models with specialized algorithmic tools in the route planning field to construct a unified intelligent agent for travel planning is not only technically feasible but also practically necessary.

[0011] These limitations indicate that systems with intelligent backend algorithms alone can no longer meet user needs. Building intelligent agents with natural language interaction, deep reasoning, and autonomous task execution capabilities is an inevitable direction to overcome existing bottlenecks. Summary of the Invention

[0012] The technical problem addressed in this application is: how to provide a method for recommending intelligent urban travel routes that has natural language interaction capabilities, can perform dynamic deep reasoning, and meets personalized needs.

[0013] This application provides a method for recommending intelligent urban travel routes, the method comprising:

[0014] Based on the analysis of users' travel demand instructions using a large language model, structured current travel elements are obtained;

[0015] Based on the current travel elements, multimodal planning is performed to obtain several candidate paths;

[0016] Based on the user's historical travel information, several candidate routes are sorted and filtered to obtain a recommended route.

[0017] Optionally, based on a large language model, the user's travel demand instructions are analyzed to obtain structured current travel elements, including:

[0018] The travel demand instructions are understood using a dynamic example enhancement algorithm based on a large language model, resulting in a structured preliminary analysis.

[0019] The preliminary analysis results are evaluated and revised based on the reflection mechanism of the large language model to obtain the final current travel elements.

[0020] Optionally, the preliminary analysis results are evaluated and revised based on a reflection mechanism of a large language model to obtain the final current travel elements, including:

[0021] When a multi-turn dialogue is detected, the travel request instruction is divided into contextual semantics and input;

[0022] An evaluation value is generated based on the contextual semantics, the current input, and the preliminary analysis results;

[0023] When the assessment value is reasonable, the preliminary analysis results will be used as the final current travel factors;

[0024] When the assessment value is unreasonable, the travel demand instruction is reinterpreted using a large language model until a preliminary analysis result with a reasonable assessment value is obtained.

[0025] Optionally, multimodal planning is performed based on the current travel elements to obtain several candidate paths, including:

[0026] The current travel elements are processed using a large language model and preset tools to obtain a preliminary planned route;

[0027] The preliminary planned path is evaluated and revised based on the reflection mechanism of the large language model to obtain the final candidate path.

[0028] Optionally, a large language model is used to call preset tools to process the current travel elements and derive a preliminary planned route, including:

[0029] The dynamic example enhancement algorithm of the large language model is used to perform semantic understanding on the current travel elements, retrieve historical similar examples based on the semantic understanding results, and generate the tool execution order based on the retrieval results;

[0030] The preset tools are invoked and executed sequentially according to the work execution order to obtain the preliminary planned path.

[0031] Optionally, the preliminary planned path is evaluated and revised based on a reflection mechanism of a large language model to obtain the final candidate path, including:

[0032] When a multi-turn dialogue is detected, the current travel element is divided into contextual semantics and current input;

[0033] Generate a path evaluation value based on contextual semantics, current input, preliminary planned path, output of each preset tool module, and generated response;

[0034] When the path evaluation value is qualified, the preliminary planned path will be used as a candidate path;

[0035] When the path evaluation value is unqualified, the large language model is used again to call the preset tool to process the current travel elements until a preliminary planned path with a qualified path evaluation value is obtained.

[0036] Optionally, the historical travel information includes user profile information and historical preference information. Based on the user's historical travel information, several candidate routes are sorted and filtered to obtain recommended routes, including:

[0037] The dynamic example enhancement algorithm using a large language model generates a sorting strategy and tool invocation order based on the user profile information, the historical preference information, and several candidate paths;

[0038] The path sorting tool is invoked and executed sequentially according to the tool invocation order to obtain the recommended path.

[0039] This application also provides an urban intelligent travel route recommendation device, the device comprising:

[0040] The demand understanding module is configured to: analyze the user's travel demand instructions based on a large language model to obtain structured current travel elements;

[0041] The planning generation module is configured to perform multimodal planning processing based on the current travel elements to obtain several candidate paths;

[0042] The personalized recommendation module is configured to sort and filter several candidate routes based on the user's historical travel information to obtain recommended routes.

[0043] This application also provides a computer-readable storage medium storing an urban intelligent travel route recommendation program, which, when executed by a processor, implements the above-described urban intelligent travel route recommendation method.

[0044] This application also provides a computer device, which includes a computer-readable storage medium, a processor, and an urban intelligent travel route recommendation program stored in the computer-readable storage medium. When the urban intelligent travel route recommendation program is executed by the processor, it implements the above-described urban intelligent travel route recommendation method.

[0045] The urban intelligent travel route recommendation method provided in this application has the following technical effects:

[0046] The recommended routes obtained by this method not only meet current travel needs and personalized preferences, but also have strong interactive and reasoning capabilities. Attached Figure Description

[0047] Figure 1 This is a flowchart of an urban smart travel route recommendation method according to one or more embodiments.

[0048] Figure 2 This is a schematic diagram of the user interaction process of an urban intelligent travel route recommendation method according to one or more embodiments.

[0049] Figure 3 This is a schematic block diagram of an urban intelligent travel route recommendation device according to one or more embodiments.

[0050] Figure 4 This is an interactive schematic diagram of the various modules of an urban intelligent travel route recommendation device according to one or more embodiments.

[0051] Figure 5 This is a schematic diagram of a computer device according to one or more embodiments. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0053] Before describing the various embodiments of this application in detail, the technical concept of this application is first briefly described: Current urban travel planning systems suffer from problems such as rigid interaction, insufficient reasoning ability, and limited personalization. To address these issues, this application provides an intelligent urban travel route recommendation method. The key improvement lies in using a large language model to analyze the user's travel demand instructions, obtaining structured current travel elements, then performing multimodal planning processing to obtain several candidate routes, and further filtering recommended routes based on historical travel information. This approach not only meets current travel needs but also aligns with personalized preferences, and possesses strong interactive and reasoning capabilities. The specific principles of this application's intelligent urban travel route recommendation method are described below with further examples.

[0054] Specifically, such as Figure 1 and Figure 2 As shown, the urban intelligent travel route recommendation method in this embodiment includes the following steps:

[0055] Step S10: Analyze the user's travel demand instructions based on the large language model to obtain structured current travel elements;

[0056] Step S20: Perform multimodal planning based on the current travel elements to obtain several candidate paths;

[0057] Step S30: Sort and filter several candidate routes based on the user's historical travel information to obtain a recommended route.

[0058] One of the core challenges of urban mobility planning lies in handling multi-dimensional constraints to meet the diverse needs of users in complex real-world scenarios. This embodiment divides constraints into two categories: "hard constraints" and "soft constraints." Travel demand instructions, as hard constraints, include conditions explicitly provided by the user and strictly enforceable, such as departure / arrival times, budget limits, necessary transit points, or designated modes of transportation, directly impacting route feasibility. Historical travel information, as soft constraints, includes conditions derived from users' historical behavior, long-term preferences, and individual characteristics, used to optimize personalized recommendations, such as preferences for routes with less walking, fewer transfers, or economical options.

[0059] In one or more embodiments, the user's travel demand instructions are analyzed based on a large language model to obtain structured current travel elements, including the following steps:

[0060] Step S101: Use the dynamic example enhancement algorithm of the large language model to understand the travel demand instructions and obtain structured preliminary analysis results;

[0061] Step S102: The preliminary analysis results are evaluated and corrected based on the reflection mechanism of the large language model to obtain the final current travel elements.

[0062] For example, the user's interactive voice is collected and converted into text commands as travel request commands. In other embodiments, the travel request commands can be text commands directly input by the user. The Large Language Model (LLM) processes the travel request commands using a Plan-firstExecution process: the LLM receives user input, dialogue context, and dynamic examples, completes request parsing, and structures the output into structured current travel elements. Current travel elements include origin, destination, waypoints, travel mode preferences (e.g., subway, bus, shared bicycle, etc.), time constraints, cost limits, and environmental conditions (e.g., weather, holiday traffic conditions, etc.).

[0063] Furthermore, the descriptive style of users' travel needs is highly flexible, making rule-based parsing difficult. Therefore, a Dynamic Demonstration-Augmented Parsing (DDAP) algorithm was added to the large language model to improve the LLM's ability to parse diverse natural language expressions and complex constraints. Its operational logic is as follows:

[0064] Example injection: Prepare a small number of high-quality examples in advance, and dynamically filter the examples that are most similar to the user's current intent based on the user's input of travel demand instructions.

[0065] Input-first method: First, generate a simulated user intent, such as: "I want to travel from Shenzhen Bay Port in Nanshan District to Meilin Yicun in Futian District, with a budget of no more than 20 yuan, and I want to minimize transfers, with the subway + shared bicycle combination as my first choice". Then, generate the corresponding structured parsing plan, such as: "Starting point: Nanshan District; Destination: Futian District; Travel mode preference: subway + shared bicycle; Cost limit: 20 yuan; Number of transfers: minimize".

[0066] Output-first method: First, a parsing plan is given, then corresponding natural language example inputs are generated to verify the plan's consistency with the user's intent. Figure 1 Consistency is crucial. For example, with a structured parsing plan like the one above, the generated user example might be: "I want to take a combination of subway and shared bike, costing no more than 20 yuan, and minimizing the number of transfers." Consistency checks filter out high-quality examples, improving the model's parsing accuracy. This dynamic example enhancement enables the handling of user input with multiple constraints, long sentences, or complex conditions.

[0067] Furthermore, in complex urban travel scenarios, users' natural language needs are often ambiguous, multi-constrained, and semantically implicit. Departure point, time constraints, and preference information all need to be inferred from the context by the model, potentially involving multiple rounds of dialogue. Although LLM possesses powerful semantic understanding capabilities, it may still encounter the following problems during the parsing process: (1) Missing elements: Incomplete travel goals, constraints, or preferences are not extracted. (2) Ambiguous parsing errors: Misjudging time, location, or budget information. (3) Logical conflicts: The parsed conditions contradict each other (e.g., a low budget but a taxi as the mode of transportation). Therefore, the preliminary analysis results need to be evaluated and corrected.

[0068] For example, the preliminary analysis results are evaluated and revised based on a reflection mechanism of a large language model to obtain the final current travel elements, including the following steps:

[0069] Step S103: When a multi-turn dialogue is detected, the travel request instruction is divided into contextual semantics and current input;

[0070] Step S104: Generate an evaluation value based on the context semantics, the current input, and the preliminary analysis results;

[0071] Step S105: When the evaluation value is reasonable, the preliminary analysis results will be used as the final current travel elements;

[0072] Step S106: When the evaluation value is unreasonable, the travel demand instruction is re-understood using the large language model until a preliminary analysis result with a reasonable evaluation value is obtained.

[0073] Specifically, in the t-th round of dialogue, the current context is C. t-1 The current input is x t Preliminary analysis results of large language models generating structured data based on contextual semantics. t This includes factors such as departure point, destination, budget, time window, travel preferences, and mid-journey constraints. A reflection mechanism verifies the consistency and completeness of the preliminary analysis results.

[0074] The first step is the evaluation, which assesses the preliminary analysis results from three dimensions. t Conduct an assessment:

[0075] 1. Semantic integrity: Whether it covers all elements, whether implicit or explicit, in the user input.

[0076] 2. Logical consistency: Are there any conflicts between the extracted elements?

[0077] 3. Operability: Whether the preliminary analysis results can be directly used for subsequent planning and processing steps (i.e., whether they meet the planning input format requirements).

[0078] The evaluation function is defined as: γ = reflect(x) t C t-1 ,d t In the expression γ∈positive,negative indicates whether the preliminary analysis results are reasonable.

[0079] Next comes error correction. When γ = negative, it indicates that the evaluation value is unreasonable, and targeted feedback signals will be generated, such as "No budget information detected" or "Origin location not specified." This guides the large language model to re-analyze the requirements based on the reflection results until a logically consistent and complete preliminary analysis result d is generated. t′ If the assessment value is deemed reasonable, then γ = positive. If the assessment value is unreasonable, the preliminary analysis results will be directly used in subsequent planning steps.

[0080] In one or more embodiments, multimodal planning is performed based on current travel elements to obtain several candidate paths, including:

[0081] Step S201: Use the large language model to call the preset tools to process the current travel elements and obtain the preliminary planned route;

[0082] Step S202: The preliminary planned path is evaluated and revised based on the reflection mechanism of the large language model to obtain the final candidate path.

[0083] In the planning process, candidate routes that meet hard constraints are generated based on structured current travel elements, including the route, required POIs, mode of transport, time window, and budget ceiling. To improve the real-time performance and stability of planning, a Plan-first Execution workflow is adopted: LLM receives current travel elements, dialogue context, toolset, and dynamic examples, generates a complete tool execution plan, and each sub-plan calls tools to complete specific tasks (such as route retrieval, real-time traffic query, and POI coordinate resolution). Tools are executed in the order of the plan to ensure that candidate routes meet all hard constraints.

[0084] For example, using a large language model to call preset tools to process the current travel elements and obtain a preliminary planned path includes: using the dynamic example enhancement algorithm of the large language model to perform semantic understanding on the current travel elements, retrieving historical similar examples based on the semantic understanding results, and generating a tool execution order based on the retrieval results; calling and executing preset tools sequentially according to the work execution order to obtain a preliminary planned path.

[0085] Specifically, for complex urban travel scenarios with multidimensional constraints (waypoints, budget, time, preferences, etc.), the Dynamic Example Augmentation Algorithm (DDAP) dynamically retrieves function-level step templates for similar tasks from a historical example library after semantic understanding, providing a clear and interpretable calling framework for LLM. Plan-first Execution generates a structured plan based on the template and then executes it step by step, ensuring the planning logic, reusability, and stability, while supporting implicit multi-constraint conditions (such as "fewer transfers," "budget < 20 yuan," and "shopping along the way").

[0086] As an example, during execution, after semantic understanding, the structured input is obtained, such as: {"origin":"Nantou Ancient City","destination":"Luohu Port","via":["Shopping Park"]","budget":30","time_constraint":"Arriving before 8:00 AM"}.

[0087] DDAP retrieves historical similar examples and adds them to the prompt, which is then sent to the LLM large model. The large model outputs a sequence of function templates: [find_viapoint_route(),evaluate_modes(),optimize_time(),select_final_plan()].

[0088] The Plan-first Execution then generates an execution plan: "1. Plan the route through points; 2. Evaluate the mode of transport and budget; 3. Optimize the time to ensure arrival before 8:00 AM; 4. Select the final optimal solution." These functions (tools) are called sequentially, with outputs such as: "Recommended Metro Line 1 (Nantou Ancient Town – Shopping Park – Luohu), time 46 minutes, cost 20 yuan." The final execution trajectory will be recorded back into the example library for future DDAP retrieval and continuous optimization.

[0089] A reflection mechanism was also used in the planning step. The preliminary planning path obtained in this step still has issues such as incompleteness, unmet constraints, or incorrect tool usage. For example, the preliminary planning path may not have considered budget constraints, time conflicts, or transportation availability. To enhance adaptability and error correction capabilities, a reflection mechanism was introduced into the planning step, enabling self-evaluation and dynamic correction of the path planning through a reflection-correction collaborative structure.

[0090] For example, the preliminary planned path is evaluated and revised based on the reflection mechanism of the large language model to obtain the final candidate path, including the following steps:

[0091] Step S203: When a multi-turn dialogue is detected, the current travel element is divided into contextual semantics and current input;

[0092] Step S204: Generate a path evaluation value based on the context semantics, current input, preliminary path planning, output of each preset tool module, and generated response;

[0093] Step S205: When the path evaluation value is qualified, the preliminary planned path is used as a candidate path;

[0094] Step S206: When the path evaluation value is unqualified, the large language model is used again to call the preset tool to process the current travel elements until a preliminary planned path with a qualified path evaluation value is obtained.

[0095] Specifically, the first step is evaluation. The reflection mechanism, based on the dialogue context and user input, re-evaluates the initially planned path p. t Output of the tool module t and generate response y t Perform a consistency and constraint rationality assessment: γ = reflect(x)t C t-1 ,p t ,o t ,y t If the route does not meet hard constraints such as budget, number of transfers, or time, or if there are abnormalities in the tool calling logic (such as duplicate calculations or omission of transportation modes), a negative feedback signal γ = negative is returned, indicating that the route evaluation value is unqualified.

[0096] Next comes error correction. When γ = negative, meaning the path evaluation value is unacceptable, the reflection mechanism feeds back error information (such as "budget exceeded" or "missing intermediate stations") to the planning processing module of the large language model, guiding it to regenerate the preliminary planned path and execute the Plan-first process again until the output meets the constraints.

[0097] When γ = positive, it indicates that the path evaluation value is reasonable, and the preliminary planned path is directly output to subsequent processing steps. This reflective mechanism enables the planning and processing module to have self-diagnosis and self-correction capabilities, effectively reducing planning deviations in LLM in multi-constraint travel scenarios and improving the feasibility and stability of the solution.

[0098] In one or more embodiments, historical travel information includes user profile information and historical preference information. Based on the user's historical travel information, several candidate routes are sorted and filtered to obtain a recommended route, including the following steps:

[0099] Step S301: Utilize the dynamic example enhancement algorithm of the large language model to generate a sorting strategy and tool invocation order based on user profile information, historical preference information, and several candidate paths;

[0100] Step S302: Call and execute the path sorting tool in the order of tool calls to obtain the recommended path.

[0101] For example, the recommendation screening step is responsible for personalized ranking and soft-constraint optimization of candidate routes, comprehensively considering both long-term and short-term user profiles, and achieving a personalized balance from multiple dimensions such as walking distance, number of transfers, landscape preferences, and cost-effectiveness. When generating recommended routes, LLM receives a set of candidate routes, user profiles, and contextual information, and dynamically retrieves historical successful examples through DDAP to extract function-level step templates and ranking strategies, thereby automatically defining the ranking methods and tool call order (such as weighted scoring, multi-dimensional comprehensive evaluation, preference matching, etc.).

[0102] For example, for the same route, the sorting can be dynamically adjusted according to user preferences: if users prefer efficient travel, routes with "few transfers and short travel time" will be recommended first; if users value travel experience, routes with "beautiful scenery but more walking" will be given higher weight.

[0103] Similar to the requirements understanding and planning steps, the recommendation filtering step generates a tool invocation plan based on the step template provided by DDAP, progressively outputting the final personalized recommendation solution. Intermediate results are temporarily stored in the memory module. A reflection mechanism is also used to achieve self-correction and strategy optimization. After execution, the sorting trajectory and feedback results are written to the example library to continuously optimize the dynamic recommendation capabilities for subsequent tasks.

[0104] like Figure 3 and Figure 4 As shown in Embodiment 2, an urban intelligent travel route recommendation device is also disclosed. The device includes a demand understanding module 100, a planning generation module 200, and a personalized recommendation module 300. The demand understanding module 100 is configured to analyze the user's travel demand instructions based on a large language model to obtain structured current travel elements. The planning generation module 200 is configured to perform multimodal planning processing based on the current travel elements to obtain several candidate routes. The personalized recommendation module 300 is configured to sort and filter several candidate routes based on the user's historical travel information to obtain recommended routes.

[0105] The urban intelligent travel route recommendation device is an intelligent agent system with a large language model as its core controller. The data processing procedures of the demand understanding module 100, the planning generation module 200, and the personalized recommendation module 300 can be referred to in detail in steps S10, S20, and S30 of the previous embodiment.

[0106] Furthermore, the urban intelligent travel route recommendation device also includes a tool usage module 400, which includes three types of tools: (1) Information acquisition tool: obtains real-time traffic conditions (such as congestion, bus delays), static traffic data (subway stations, pedestrian walkways) and event information (construction, event control) through API or structured interfaces, and parses POI / place names to obtain precise coordinates, providing constraints for route retrieval and ranking. (2) Route retrieval tool: generates a set of candidate routes. Multimodal travel (subway, bus, cycling) uses multimodal planning algorithms to generate candidate routes, achieving seamless connection of transfer points, time and mode of transportation; driving or walking routes can be combined with relevant algorithms to calculate the fastest or shortest route. (3) Route ranking tool: combines user preferences, real-time data and historical behavior to perform multi-objective ranking of candidate routes. Personalized ranking can capture long-term preferences (such as saving time, low carbon, comfort), and multi-objective optimization can comprehensively analyze time, cost, comfort and low carbon, and generate a final recommendation scheme that takes into account efficiency, economy and user experience.

[0107] Furthermore, the urban intelligent travel route recommendation device also includes a memory module 500. The memory module 500 maintains user profiles and a long-term knowledge base, records historical travel behavior, traffic preferences, and frequently used routes, captures implicit preferences (such as walking distance, number of transfers, and preferences for economical or scenic routes), supports multi-round interaction and personalized optimization, and provides contextual information for the recommendation and planning modules, enabling the system to gradually adapt to user needs as it is used. The memory module 500 consists of two parts: a short-term memory submodule and a long-term memory submodule.

[0108] The short-term memory submodule records the user's current travel request, dialogue context, immediate constraints, and tool call history, including information such as origin, destination, waypoints, travel mode preferences, time limits, and cost limits. Short-term memory provides the basis for candidate path generation, ranking, and reflection mechanisms, enabling the system to respond quickly in multi-round interactions and make real-time adjustments and optimizations to the solutions. Each tool call generates a record, represented as a triple f. k i k ,o k , where f k Let i represent the name of the k-th tool. k and o k These represent the input and output of the tool during execution, such as the number of remaining candidate paths and runtime errors. This record supports the evaluation process in the reflexive mechanism, equivalent to the o in the reflexive function reflect(·). k This provides a reference for the system to judge and optimize the generated scheme.

[0109] The long-term memory submodule is used to accumulate users' historical travel data and preferences, building a complete user profile and supporting personalized route ranking and multi-objective optimization. It records users' stable preferences and long-term behavioral patterns, such as elderly users preferring routes with less walking and fewer transfers; pregnant women or people with mobility impairments avoiding cycling or stairs. In addition, long-term memory includes important fixed-location information, such as home address, workplace, frequently visited schools, or commonly used points of interest (POIs), providing spatial constraints for route planning. It also combines historical behavioral data to predict user preference trends, enabling RouteAgent to generate travel plans that meet both immediate constraints and long-term habits and personalized needs.

[0110] Embodiment 3 of this application also discloses a computer-readable storage medium storing an urban intelligent travel route recommendation program, which, when executed by a processor, implements the above-described urban intelligent travel route recommendation method.

[0111] This fourth embodiment also discloses a computer device, at the hardware level, such as... Figure 5As shown, the computer device includes a processor 12, an internal bus 13, a network interface 14, and a computer-readable storage medium 11. The processor 12 reads the corresponding computer program from the computer-readable storage medium and runs it, forming a request processing device at the logical level. Of course, in addition to the software implementation, one or more embodiments of this specification do not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices. The computer-readable storage medium 11 stores a smart city travel route recommendation program, which, when executed by the processor, implements the above-described smart city travel route recommendation method.

[0112] Computer-readable storage media include both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer-readable storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage, quantum memory, graphene-based storage media or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0113] The specific embodiments of this application have been described in detail above. Although some embodiments have been shown and described, those skilled in the art should understand that modifications and improvements can be made to these embodiments without departing from the principles and spirit of this application as defined by the claims and their equivalents, and such modifications and improvements should also be within the protection scope of this application.

Claims

1. A method for recommending an urban intelligent travel path, characterized in that, The method comprises: analyzing the travel demand instruction of the user based on a large language model to obtain structured current travel elements; performing multi-modal planning processing according to the current travel elements to obtain a plurality of candidate paths; sorting and screening the plurality of candidate paths according to the historical travel information of the user to obtain a recommended path. 2.The urban intelligent travel path recommendation method of claim 1, wherein, The method comprises: analyzing the travel demand instruction of the user based on a large language model to obtain structured current travel elements, comprising: understanding the travel demand instruction by using a dynamic example enhancement algorithm of the large language model to obtain a structured preliminary analysis result; 3.The urban intelligent travel path recommendation method of claim 2, wherein, evaluating and correcting the preliminary analysis result based on a reflection mechanism of the large language model to obtain the final current travel elements. The method comprises: when it is identified that there is multi-round dialogue, dividing the travel demand instruction into context semantics and input; generating an evaluation value according to the context semantics, the current input and the preliminary analysis result; when the evaluation value is reasonable, taking the preliminary analysis result as the final current travel elements; 4.The urban intelligent travel path recommendation method of claim 1, wherein, when the evaluation value is not reasonable, re-understanding the travel demand instruction by using the large language model until a preliminary analysis result with a reasonable evaluation value is obtained. The method comprises: processing the current travel elements by using the large language model to call preset tools to obtain a preliminary planning path; 5.The urban intelligent travel path recommendation method of claim 4, wherein, evaluating and correcting the preliminary planning path based on a reflection mechanism of the large language model to obtain the final candidate path. The method comprises: understanding the semantics of the current travel elements by using a dynamic example enhancement algorithm of the large language model, retrieving historical similar examples according to the result of semantic understanding, and generating a tool execution sequence according to the retrieval result; 6.The urban intelligent travel path recommendation method of claim 5, wherein, calling preset tools in sequence according to the work execution sequence and executing to obtain a preliminary planning path. The method comprises: when it is identified that there is multi-round dialogue, dividing the current travel elements into context semantics and current input; generating a path evaluation value according to the context semantics, the current input, the preliminary planning path, the output of each preset tool module and the generated response generation path; when the path evaluation value is qualified, taking the preliminary planning path as the candidate path; 7.The urban intelligent travel path recommendation method of claim 6, wherein, when the path evaluation value is unqualified, re-processing the current travel elements by using the large language model to call preset tools until a preliminary planning path with a qualified path evaluation value is obtained. The historical travel information comprises user portrait information and historical preference information, and the method comprises: generating a sorting strategy and a tool calling sequence according to the user portrait information, the historical preference information and the plurality of candidate paths by using a dynamic example enhancement algorithm of the large language model. The path ranking tool is called in sequence according to the tool calling sequence and executed to obtain a recommended path.

8. An urban intelligent travel path recommendation device, characterized by, The device comprises: The demand understanding module is configured to analyze the travel demand instruction of the user based on a large language model to obtain a structured current travel element; The planning generation module is configured to perform multi-modal planning processing according to the current travel element to obtain a plurality of candidate paths; The personalized recommendation module is configured to sort and filter the plurality of candidate paths according to historical travel information of the user to obtain a recommended path.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a city intelligent travel path recommendation program, and the city intelligent travel path recommendation program, when executed by the processor, implements the city intelligent travel path recommendation method in any one of claims 1 to 7.

10. A computer device, comprising: The computer device comprises a computer readable storage medium, a processor and a city intelligent travel path recommendation program stored in the computer readable storage medium, and the city intelligent travel path recommendation program, when executed by the processor, implements the city intelligent travel path recommendation method in any one of claims 1 to 7.

Images