A parallel scheduling and control method for large model agents in vehicle-mounted dialogue systems

Abstract

The present application relates to a method and system for parallel scheduling and control of a large-model intelligent agent in an in-vehicle dialogue system, and relates to the technical field of automobile interaction. The method includes: obtaining voice input from a user in the vehicle, parsing the voice input to obtain multiple user intentions and skills corresponding to the multiple user intentions; initiating calls to multiple intelligent agents based on the multiple user intentions and the skills corresponding to the multiple user intentions; inputting information of the multiple intelligent agent modules that initiate the calls into the large model to obtain output results of the large model output; the output results include intelligent agents that need to be actually called; and, based on the output results, controlling the actually called intelligent agents to work.

Description

Technical Field

[0001] The present application relates to the technical field of automobile interaction, and in particular to a method and system for parallel scheduling and control of a large model intelligent agent in an in-vehicle dialogue system. Background Art

[0002] At present, the semantics of vehicle cockpit dialogue systems are all in the form of instructions. When users give clear instructions, the in-vehicle dialogue system performs semantic recognition and takes the next step based on the instructions. Current solutions generally have insufficient generalization capabilities. For example, they can recognize the instruction to turn on the air conditioner, but cannot recognize the generalized instruction to adjust the air conditioner to a suitable temperature.

[0003] To solve the generalization problem, large model capabilities are widely used in dialogue systems. With the help of large model reasoning and context understanding capabilities, traditional prediction skills can be further improved to identify the specific intentions of users, which has a very good effect. The introduction of large model capabilities has improved the generalization ability of speech, making the conversation between people and cars more and more natural. However, it also brings new problems. The introduction of large model technology has made users have higher requirements for the generalization system, expecting the system to better understand the intentions in the conversation. Since there are naturally multi-intention scenarios in human conversational speech, the current command-based accurate classification method in the in-vehicle dialogue system is no longer applicable to the broad generalized intentions after the support of large models, affecting the user experience.

[0004] Therefore, how to achieve smarter human-vehicle interaction that better meets user needs has become a problem that needs to be solved urgently. Summary of the invention

[0005] In order to at least partially solve the above technical problems, the present application provides a method and system for parallel scheduling and control of a large model intelligent agent in an in-vehicle dialogue system.

[0006] In a first aspect, the present application provides a method for parallel scheduling and control of a large model intelligent agent in an in-vehicle dialogue system, which adopts the following technical solution.

[0007] A method for parallel scheduling and control of a large model agent in a vehicle-mounted dialogue system, comprising:

[0008] Acquire a voice input from a user in the vehicle, and analyze the voice input to obtain multiple user intentions and skills corresponding to the multiple user intentions;

[0009] Initiate calls to multiple agents based on the multiple user intentions and the skills corresponding to the multiple user intentions;

[0010] Inputting the information of multiple agent modules that initiate the call into the large model to obtain the output result of the large model; the output result includes the agent that needs to be actually called; and

[0011] The actually called intelligent agent is controlled to perform work based on the output result.

[0012] Optionally, a voice input of a user in the vehicle is obtained, and the voice input is parsed to obtain multiple user intentions, including:

[0013] Parsing the voice input of the user in the car to obtain a user input x;

[0014] The user input x is inferred using a large model to obtain the output probability corresponding to each agent; wherein the large model is provided with an agent set and an agent skill weight; each element in the agent set corresponds to an intention; the agent skill weight includes a preset weight for each agent; the weight is the importance of the agent in the reasoning process;

[0015] For each agent, determine whether its output probability is greater than the preset probability threshold; if so, it is determined to meet the conditions;

[0016] Output all agents that meet the conditions as multiple user intentions obtained by parsing.

[0017] Optionally, the information of the multiple agent modules that initiate the call is input into the large model to obtain the output results of the large model, including:

[0018] Determine the response delay weight:

[0019] Determine agent priority weights Among them, each agent S i Corresponding to a static priority weight

[0020] Determine the weight of user personalized preferences Used to characterize user preferences for different skills;

[0021] Determine vehicle restriction status weights

[0022] The arbitration function A is determined based on the response delay weight, the agent priority weight, the user personalized preference weight and the vehicle restriction state weight; wherein the formula of the arbitration function A is: P i It is the agent S i The output probability distribution of

[0023] Select the skill S with the largest value of Arbitration Function A k As the output of the skill.

[0024] Optionally, the method further comprises: dynamically updating the weight of the agent according to user feedback and preferences;

[0025] Based on user feedback and preferences, the weights of the agent are dynamically updated, including:

[0026] Set the weight vector w and update function U; where the weight vector w contains each agent S i The update function U is used to dynamically adjust the weight;

[0027] Set the feedback vector r; the feedback vector r contains the user's feedback information on each agent's output;

[0028] The update function U takes the feedback vector r and the current weight vector w as input to output a new weight vector w'; where w' = w + α*r; where α is the learning rate, which is used to control the update speed, 0 < α < < 1;

[0029] The updated weight vector w' will be used in the next agent arbitration process. During the arbitration process, each agent S i The output probability distribution P i Will be weighted according to the corresponding weight; n is the total number of agents.

[0030] Optionally, the method further includes: collecting current vehicle status information, and analyzing the vehicle status information through a large model to obtain a current situation;

[0031] The agent actually invoked is adjusted according to the current situation.

[0032] Optionally, current vehicle status information is collected, and the vehicle status information is analyzed by a large model to obtain the current situation. The specific steps include:

[0033] Performing a preliminary analysis on the vehicle status information to identify at least one key feature for current situation judgment;

[0034] Set a starting reasoning path and feature weights based on the branching logic of the decision tree;

[0035] Perform reasoning based on the current reasoning path and adjust the path based on output feedback and adjust feature weights through backpropagation during the path adjustment process;

[0036] After multiple reasoning paths, a convergent reasoning path is obtained, and the current situation is obtained based on the convergent reasoning path.

[0037] Optionally, initiating calls to multiple agents, including: initiating parallel calls to multiple agents.

[0038] Optionally, the vehicle restriction status includes: vehicle speed, gear information, road conditions and vehicle static information.

[0039] Optionally, the intelligent agent includes: weather, navigation, consultation, encyclopedia, car manual and graphic diagram.

[0040] In the second aspect, the present application provides a method for parallel scheduling and control of a large model intelligent agent in a vehicle-mounted dialogue system, which adopts the following technical solution.

[0041] A large model agent parallel scheduling and control system in a vehicle-mounted dialogue system, characterized by comprising:

[0042] A multi-channel classification module is used to: obtain a voice input from a user in the vehicle, and analyze the voice input to obtain multiple user intentions and skills corresponding to the multiple user intentions;

[0043] An agent module, including multiple agents;

[0044] A multi-way calling module, used to: initiate calls to multiple agents based on the multiple user intentions and the skills corresponding to the multiple user intentions;

[0045] A multi-path judgment module is used to: input the information of multiple intelligent agent modules that initiate the call into the large model to obtain the output result of the large model; the output result includes the intelligent agent that needs to be actually called;

[0046] The execution module is used to control the actually called intelligent agent to work based on the output result. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flow chart of a method for parallel scheduling and control of a large model agent in a vehicle-mounted dialogue system according to an embodiment of the present application;

[0048] Figure 2 This is a system block diagram of a method for parallel scheduling and control of a large model agent in a vehicle-mounted dialogue system according to an embodiment of the present application;

[0049] In the figure, 201 is a multi-way classification module; 202 is an intelligent agent module; 203 is a multi-way calling module; 204 is a multi-way judgment module; 205 is an execution module. DETAILED DESCRIPTION

[0050] Big model capabilities are widely used in dialogue systems. With the help of the reasoning and context understanding capabilities of big models, traditional prediction skills can be further improved to identify the specific intentions of users, which has a very good effect. The introduction of big model capabilities has improved the generalization ability of speech, making the dialogue between people and cars more and more natural. However, it also brings new problems. The introduction of big model technology has made users have higher requirements for the generalized system, and they expect the system to better understand the intentions in the dialogue. Since there are naturally multiple intention scenarios in human dialogue speech, the current command-based accurate classification method in the in-vehicle dialogue system can no longer be applied to the broad generalized intentions after the big model is added, affecting the user experience. For example, when a user in the dialogue system requests: What is the weather like in Beijing during this period? Follow-up surveys of actual users show the following possible intentions:

[0051] Today's weather forecast for Beijing, the corresponding skill is Weather Report;

[0052] The overall change of climate in the recent long period of time, the corresponding skill is the overall weather query;

[0053] Based on geographical location and historical conditions, the overall weather conditions in Beijing during this period, the corresponding skill is encyclopedic knowledge;

[0054] Since the current in-vehicle dialogue system understands human language in a single classification, it specifically selects a single skill and slot to achieve a closed-loop function. However, the user's specific intention is related to the vehicle status, geographic location, and mood. A single skill classification cannot meet the user's situational needs.

[0055] In order to solve the above problems, the present application proposes a method and system for parallel scheduling and control of a large model intelligent agent in a vehicle-mounted dialogue system.

[0056] The following is combined with Figure 1-2 The present application is further described with specific embodiments:

[0057] First of all, it should be noted that in the description of this application, if the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inside", "outside" and other directional words appear, the orientation or position relationship indicated is based on the orientation or position relationship shown in the drawings, which is only for the convenience of description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation of this application; in addition, if the terms "first", "second", "third" and other numerical quantifiers appear, they are only used for descriptive purposes and cannot be understood as indicating or implying relative importance. In addition, in this application, unless otherwise clearly specified and limited, if the terms "installed", "connected", and "connected" appear, they should be understood in a broad sense, for example, it can be a fixed connection, or a detachable connection, a limited connection such as an interference fit, a transition fit, or an integral connection; it can be directly connected or indirectly connected through an intermediate medium; therefore, for ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to the specific circumstances.

[0058] The present application embodiment discloses a method for parallel scheduling and control of a large model agent in a vehicle-mounted dialogue system. Figure 1 As an implementation method of a large model agent parallel scheduling and control method in a vehicle-mounted dialogue system, a large model agent parallel scheduling and control method in a vehicle-mounted dialogue system includes:

[0059] Step S101: Acquire a voice input from a user in a vehicle, and analyze the voice input to obtain multiple user intentions and skills corresponding to the multiple user intentions.

[0060] Step S102: Initiate calls to multiple agents based on the multiple user intentions and the skills corresponding to the multiple user intentions.

[0061] Step S103: Input the information of the multiple agent modules that initiate the call into the large model to obtain the output result of the large model; the output result includes the agent that needs to be actually called. And,

[0062] Step S104: Control the actually called intelligent agent to work based on the output result.

[0063] Specifically, the user's voice input is parsed to identify multiple user intentions and their corresponding skill requirements, not just one; based on the identified multiple user intentions and skill requirements, the corresponding agent is called. This intention-based calling mechanism can avoid unnecessary agent calls, reduce resource waste, and improve the overall performance of the system. The large model is used to further filter out the agents that really need to be called. This step helps to filter out redundant agent call requests. The actual called agent is controlled to work according to the output results of the large model, which ensures effective collaboration between the agents and enables the system to complete user instructions more smoothly.

[0064] As a specific implementation method of a large model agent parallel scheduling and control method in a vehicle-mounted dialogue system, a voice input of a user in the vehicle is obtained, and the voice input is parsed to obtain multiple user intentions, including:

[0065] Parsing the voice input of the user in the car to obtain a user input x;

[0066] The user input x is inferred using a large model to obtain the output probability corresponding to each agent; wherein the large model is provided with an agent set and an agent skill weight; each element in the agent set corresponds to an intention; the agent skill weight includes a preset weight for each agent; the weight is the importance of the agent in the reasoning process;

[0067] For each agent, determine whether its output probability is greater than the preset probability threshold; if so, it is determined to meet the conditions;

[0068] Output all agents that meet the conditions as multiple user intentions obtained by parsing.

[0069] Specifically, the voice input of the user in the car is parsed to obtain the user input x, and the original voice data is accurately converted into a text form that can be used for subsequent processing; the user input x is inferred using a large model to obtain the output probability corresponding to each agent, and the relevance of each agent to the user's intention is evaluated, thereby improving the accuracy of intent recognition. Each agent has a specific weight, which helps the model better understand which agents are more important for a specific task, thereby giving higher priority when identifying user intentions. For each agent, it is determined whether its output probability is greater than a preset probability threshold, so that only agents with high probability matching will be selected, thereby reducing the situation of mismatching and improving the system's response quality and user experience. The system outputs all agents that meet the conditions as multiple user intentions obtained by parsing, and screens out a group of agents that are most likely to meet the user's intention, which enables the system to respond to user requests more accurately.

[0070] As a specific implementation method of a large model agent parallel scheduling and control method in a vehicle-mounted dialogue system, information of multiple agent modules that initiate calls is input into the large model to obtain the output result of the large model, including:

[0071] Determine the response delay weight:

[0072] Determine agent priority weights Among them, each agent S i Corresponding to a static priority weight

[0073] Determine the weight of user personalized preferences Used to characterize user preferences for different skills;

[0074] Determine vehicle restriction status weights

[0075] The arbitration function A is determined based on the response delay weight, the agent priority weight, the user personalized preference weight and the vehicle restriction state weight; wherein the formula of the arbitration function A is: P i It is the agent S i The output probability distribution of

[0076] Select the skill S with the largest value of Arbitration Function A k As the output of the skill.

[0077] Specifically, by considering the response time of the agent, those agents with fast response speed can be given priority, thereby speeding up the response speed of the entire system; each agent has a static priority weight, and some key or commonly used functions will be given priority to ensure that important functions are processed first and improve user experience. Considering the personalized needs of users, the system can adjust the selection of agents according to the user's historical behavior and preferences, so that the service is more in line with the user's habits and enhances user satisfaction. The state of the vehicle (such as driving status, safety requirements, etc.) will affect the available agents and services. By considering these restrictions, the system can make more reasonable and safe service decisions. By comprehensively considering the above factors, the arbitration function A can comprehensively evaluate the applicability of each agent, so as to select the most suitable skills for the current situation. The skill with the largest value of arbitration function A is selected as the output result of the skill, ensuring that the best skill is selected in the current situation, maximizing the satisfaction of user needs while also taking into account factors such as the response speed and safety of the system.

[0078] As a specific implementation of a method for parallel scheduling and control of a large model agent in a vehicle-mounted dialogue system, the method further includes: dynamically updating the weight of the agent according to user feedback and preferences;

[0079] Based on user feedback and preferences, the weights of the agent are dynamically updated, including:

[0080] Set the weight vector w and update function U; where the weight vector w contains each agent S i The update function U is used to dynamically adjust the weight;

[0081] Set the feedback vector r; the feedback vector r contains the user's feedback information on each agent's output;

[0082] The update function U takes the feedback vector r and the current weight vector w as input to output a new weight vector w'; where w' = w + α*r; where α is the learning rate, which is used to control the update speed, 0 < α <<1;

[0083] The updated weight vector w' will be used in the next agent arbitration process. During the arbitration process, each agent S i The output probability distribution P i Will be weighted according to the corresponding weight; n is the total number of agents.

[0084] Specifically, by collecting user feedback and preferences, the system can continuously adjust the weights of the agents to make them more in line with the user's personal preferences. The weight vector w contains the weights of each agent, and the update function U is used to dynamically adjust these weights based on user feedback. The feedback vector r contains the user's feedback information on the output of each agent, which is used to guide the update of the weights. By utilizing user feedback, the system can better understand the user's needs and make more accurate service decisions based on this. By using the learning rate α to control the speed of weight updates, it can ensure that the system can quickly adapt to user needs while avoiding instability caused by over-adjustment. During the arbitration process, the output probability distribution of each agent is weighted according to the corresponding weight. The system will be more inclined to select those agents that have been verified by users to perform better, thereby providing higher quality services.

[0085] As one implementation of a method for parallel scheduling and control of a large model agent in a vehicle-mounted dialogue system, the method further includes: collecting current vehicle state information, analyzing the vehicle state information through a large model to obtain a current situation;

[0086] The agent actually invoked is adjusted according to the current situation.

[0087] As one implementation of a large-model intelligent agent parallel scheduling and control method in a vehicle-mounted dialogue system, current vehicle state information is collected, and the vehicle state information is analyzed by a large model to obtain the current situation. The specific steps include:

[0088] Performing a preliminary analysis on the vehicle status information to identify at least one key feature for current situation judgment;

[0089] Set a starting reasoning path and feature weights based on the branching logic of the decision tree;

[0090] Perform reasoning based on the current reasoning path and adjust the path based on output feedback and adjust feature weights through backpropagation during the path adjustment process;

[0091] After multiple reasoning paths, a convergent reasoning path is obtained, and the current situation is obtained based on the convergent reasoning path.

[0092] Specifically, by conducting a preliminary analysis of the key features of the vehicle status information, the system can quickly capture important clues to the current situation; through the branching logic of the decision tree, the system can select the appropriate reasoning path based on the current key features and set the corresponding weights according to the importance of the features; reasoning is performed based on the current reasoning path and the path is adjusted based on the output feedback and the feature weights are adjusted by backpropagation during the path adjustment process. The system dynamically adjusts the reasoning path based on the feedback of the reasoning results and adjusts the feature weights by backpropagation, continuously optimizing the process of situation recognition and improving accuracy. Through multiple iterative reasoning until the path converges, the system can ensure the stability of the situation recognition results and reduce the errors caused by a single reasoning path.

[0093] As one implementation method of a large-model intelligent agent parallel scheduling and control method in a vehicle-mounted dialogue system, initiating calls to multiple intelligent agents includes: initiating parallel calls to multiple intelligent agents.

[0094] As one implementation method of a large-model intelligent agent parallel scheduling and control method in a vehicle-mounted dialogue system, the vehicle restriction status includes: vehicle speed, gear information, road conditions and vehicle static information.

[0095] As one implementation method of a large-model intelligent agent parallel scheduling and control method in a vehicle-mounted dialogue system, the intelligent agent includes: weather, navigation, consultation, encyclopedia, vehicle manual and cultural map.

[0096] The present application also discloses a large model agent parallel scheduling and control system in a vehicle-mounted dialogue system, including:

[0097] The multi-channel classification module 201 is used to obtain the voice input of the user in the car, analyze the voice input to obtain multiple user intentions and skills corresponding to the multiple user intentions;

[0098] The agent module 202 includes a plurality of agents;

[0099] A multi-way calling module 203 is used to: initiate a call to multiple agents based on the multiple user intentions and the skills corresponding to the multiple user intentions;

[0100] The multi-path judgment module 204 is used to: input the information of the multiple agent modules that initiate the call into the large model to obtain the output result of the large model; the output result includes the agent that needs to be actually called;

[0101] The execution module 205 is used to control the actually called intelligent agent to work based on the output result.

[0102] It should be noted that the above embodiments are only used to illustrate the present application and are not intended to limit the technical solutions described in the present application. Although the present application has been described in detail in this specification with reference to the above embodiments, a person of ordinary skill in the art should understand that a person of ordinary skill in the art can still modify or make equivalent substitutions to the present application, and all technical solutions and improvements thereof that do not depart from the spirit and scope of the present application should be included in the scope of the claims of the present application.

Claims

1. A method for parallel scheduling and control of large model agents in a vehicle-mounted dialogue system, characterized in that: include: Acquire a voice input from a user in the vehicle, and analyze the voice input to obtain multiple user intentions and skills corresponding to the multiple user intentions; Initiate calls to multiple agents based on the multiple user intentions and the skills corresponding to the multiple user intentions; Input the information of multiple agent modules that initiate the call into the large model to obtain the output results of the large model; The output result includes the intelligent agent that needs to be actually called; as well as, Controlling the actually called intelligent agent to work based on the output result; The step of inputting the information of the multiple agent modules that initiate the call into the large model to obtain the output result of the large model includes: Determine the response delay weight: Determine agent priority weights Among them, each agent S i Corresponding to a static priority weight Determine the weight of user personalized preferences Used to characterize user preferences for different skills; Determine vehicle restriction status weights The arbitration function A is determined based on the response delay weight, the agent priority weight, the user personalized preference weight and the vehicle restriction state weight; wherein the formula of the arbitration function A is: P i It is the agent S i The output probability distribution of Select the skill S with the largest value of Arbitration Function A k As the output of the skill.

2. The method for parallel scheduling and control of a large model agent in a vehicle-mounted dialogue system according to claim 1, characterized in that: Acquire the voice input of the user in the car, and parse the voice input to obtain multiple user intentions, including: Parsing the voice input of the user in the car to obtain a user input x; The user input x is inferred using a large model to obtain the output probability corresponding to each agent; wherein the large model is provided with an agent set and an agent skill weight; each element in the agent set corresponds to an intention; the agent skill weight includes a preset weight for each agent; the weight is the importance of the agent in the reasoning process; For each agent, determine whether its output probability is greater than the preset probability threshold; if so, it is determined to meet the conditions; Output all agents that meet the conditions as multiple user intentions obtained by parsing.

3. The method for parallel scheduling and control of a large model agent in a vehicle-mounted dialogue system according to claim 2, characterized in that: The method further includes: dynamically updating the weight of the agent based on user feedback and preferences; Based on user feedback and preferences, the weights of the agent are dynamically updated, including: Set the weight vector w and update function U; where the weight vector w contains each agent S i The update function U is used to dynamically adjust the weight; Set the feedback vector r; the feedback vector r contains the user's feedback information on each agent's output; The update function U takes the feedback vector r and the current weight vector w as input to output a new weight vector w'; where w' = w + α*r; where α is the learning rate, which is used to control the update speed, 0 < α < < 1; The updated weight vector w' will be used in the next agent arbitration process. During the arbitration process, each agent S i The output probability distribution P i Will be weighted according to the corresponding weight; n is the total number of agents.

4. The method for parallel scheduling and control of a large model agent in a vehicle-mounted dialogue system according to claim 3, characterized in that: The method further includes: collecting current vehicle status information, and analyzing the vehicle status information through a large model to obtain a current situation; The agent actually invoked is adjusted according to the current situation.

5. The method for parallel scheduling and control of a large model agent in a vehicle-mounted dialogue system according to claim 4, characterized in that: Collect the current vehicle status information, and analyze the vehicle status information through the big model to obtain the current situation. The specific steps include: Performing a preliminary analysis on the vehicle status information to identify at least one key feature for current situation judgment; Set a starting reasoning path and feature weights based on the branching logic of the decision tree; Perform reasoning based on the current reasoning path and adjust the path based on output feedback and adjust feature weights through backpropagation during the path adjustment process; After multiple reasoning paths, a convergent reasoning path is obtained, and the current situation is obtained based on the convergent reasoning path.

6. The method for parallel scheduling and control of a large model agent in a vehicle-mounted dialogue system according to claim 5, characterized in that: Initiating calls to multiple agents, including: initiating parallel calls to multiple agents.

7. The method for parallel scheduling and control of a large model agent in a vehicle-mounted dialogue system according to claim 6, characterized in that: The vehicle restriction status includes: vehicle speed, gear information, road conditions and vehicle static information.

8. The method for parallel scheduling and control of a large model agent in a vehicle-mounted dialogue system according to claim 7, characterized in that: The intelligent agents include: weather, navigation, consultation, encyclopedia, car manual and cultural map.

9. A large model agent parallel scheduling and control system in a vehicle-mounted dialogue system, characterized in that: include: A multi-channel classification module is used to obtain a voice input from a user in the vehicle, analyze the voice input to obtain multiple user intentions and skills corresponding to the multiple user intentions; An agent module, including multiple agents; A multi-way calling module, used to: initiate calls to multiple agents based on the multiple user intentions and the skills corresponding to the multiple user intentions; The multi-path judgment module is used to: input the information of the multiple intelligent agent modules that initiate the call into the large model to obtain the output result of the large model; The output result includes the intelligent agent that needs to be actually called; The step of inputting the information of the multiple agent modules that initiate the call into the large model to obtain the output result of the large model includes: Determine the response delay weight: Determine agent priority weights Among them, each agent S i Corresponding to a static priority weight Determine the weight of user personalized preferences Used to characterize user preferences for different skills; Determine vehicle restriction status weights The arbitration function A is determined based on the response delay weight, the agent priority weight, the user personalized preference weight and the vehicle restriction state weight; wherein the formula of the arbitration function A is: P i It is the agent S i The output probability distribution of Select the skill S with the largest value of Arbitration Function A k As the output of skills; The execution module is used to control the actually called intelligent agent to work based on the output result.

Images