A system and development method for testing the social ability of intelligent agents

Abstract

The present invention provides an intelligent agent social ability testing system and development method, including: a three-dimensional modeling dynamic function building subsystem, which uses three-dimensional modeling and Unreal Engine to obtain static city scenes and develop digital city dynamic functions to build a three-dimensional digital city; a social network interaction model subsystem, which builds a memory-inductive interaction intelligent agent in the digital city and a social network model; an intelligent agent access acceleration training subsystem, which selects an embodied character, connects to the digital city, and simultaneously starts an interpersonal trust algorithm expert model, and accelerates the interaction training speed of the memory-inductive interaction intelligent agent and the intelligent agent to be tested; an embodied observation training output subsystem, which synchronizes observation data to the intelligent agent to be tested and the interpersonal trust algorithm expert model, scores the feedback content of the intelligent agent to be tested, and outputs an intelligent agent social ability test report.

Description

Technical Field

[0001] The present invention relates to the technical field of digital city intelligent interaction model testing, and more specifically, to an intelligent agent social ability testing system and a development method. Background Art

[0002] Currently, intelligent agents are increasingly used in various fields, including intelligent assistants, emotional robots, personalized NPCs, etc. In these applications, the social ability of intelligent agents has become a key factor, directly affecting the user experience and the effectiveness of the system. For intelligent agents, social ability not only involves understanding and responding to human behavior, but also includes showing appropriate behavior and emotional reactions in a variety of social situations. This requires intelligent agents to have natural language processing, emotion recognition, behavior prediction and decision-making capabilities. Therefore, how to evaluate the social ability of intelligent agents has become an important research topic;

[0003] At present, the social ability of intelligent agents is mainly achieved through two methods. One is to rely on preset rules and scripts to conduct social interactions through fixed dialogue trees and behavior patterns. This type of intelligent agent is limited to existing settings, the interaction method is fixed and single, and lacks flexibility and adaptability. The other type is a dialogue system trained based on large-scale dialogue data. This type of system can achieve more natural and diverse dialogue content, but it is strongly dependent on user expression and is limited by the limited input of users. It is difficult to obtain a complete situation description for more real and natural social interaction, and thus cannot exert its real social ability. This type of system is also the main test object in the present invention. However, there is currently a lack of an automated testing system that can input all-factor information to comprehensively test the social ability of intelligent agents, which is also a key issue worthy of study; there is no existing testing system to test the social ability of intelligent agents; there is a lack of rich and diverse test scenarios that can conduct comprehensive social ability tests on intelligent agents; there is no computational model to support the quantitative evaluation of the social ability of intelligent agents. These problems remain to be solved; therefore, it is necessary to propose a system and method for testing the social ability of intelligent agents to at least partially solve the problems existing in the prior art. Summary of the invention

[0004] A series of simplified concepts are introduced in the content of the invention, which will be further described in detail in the specific implementation method section; the content of the invention of the present invention does not mean to attempt to limit the key features and essential technical features of the technical solution claimed for protection, nor does it mean to attempt to determine the scope of protection of the technical solution claimed for protection.

[0005] To at least partially solve the above problems, the present invention provides an intelligent agent social ability testing system, comprising:

[0006] The 3D modeling dynamic function construction subsystem uses 3D modeling and Unreal Engine to obtain static city scenes and develop digital city dynamic functions to build a 3D digital city;

[0007] The social network interaction model subsystem builds memory-inductive interaction agents and social network models in the digital city.

[0008] The agent accesses the accelerated training subsystem, selects an embodied character, and accesses the digital city. At the same time, the interpersonal trust algorithm expert model is started, and the interactive training speed of the memory induction interactive agent and the test agent is accelerated;

[0009] The embodied observation training output subsystem synchronizes observation data to the intelligent agent to be tested and the expert model of the interpersonal trust algorithm, scores the feedback content of the intelligent agent to be tested, and outputs a test report on the social ability of the intelligent agent.

[0010] Preferably, the 3D modeling dynamic function building subsystem includes:

[0011] The static city scene construction module builds a digital city 3D model in a 3D modeling tool and designs and builds static city scenes.

[0012] Digital city dynamic function module, using Unreal Engine to define and develop digital city dynamic functions.

[0013] Preferably, the social network interaction model subsystem includes:

[0014] The interactive agent model module builds a memory-inductive interactive agent based on a large language model;

[0015] Social network construction module, in the digital city, builds a social network model based on the social network structure;

[0016] In a digital city, a social network model is constructed based on a social network structure, including: presetting multiple types of character information, importing the preset multiple types of character information into a development program in the form of a data table, configuring a mapping relationship between the character information and the character model through the development program, and generating embodied characters that conform to the preset multiple types of character information in a corresponding area in a digital city scene; constructing a social network model based on the social network structure; the social network structure includes: class structure, family structure, regional structure and employment structure; multiple types of character information include: gender information, age information, appearance information, occupation information, address information and social relationship information.

[0017] Preferably, the intelligent agent accesses the accelerated training subsystem, including:

[0018] The embodied access model startup module, in which the intelligent agent to be tested is connected to the digital city as an embodied person, and the interpersonal trust algorithm expert model is started at the same time;

[0019] The interaction frequency acceleration training module speeds up the interaction training of the intelligent agent to be tested by increasing the interaction frequency of interaction requests with the intelligent agent to be tested.

[0020] Preferably, the embodied observation training output subsystem includes:

[0021] Synchronous embodied observation module: the test system synchronizes embodied observation data to the tested intelligent agent and the interpersonal trust algorithm expert model at a set frequency; the embodiment performs interactive response actions based on the feedback from the tested intelligent agent; and the interpersonal trust algorithm expert model gives feedback;

[0022] The training cycle scoring output module sets the number of interactive training cycles, takes the number of interactive training cycles as a node, ends the test after multiple rounds of interaction are completed, scores the feedback content of the tested intelligent agent, and outputs the intelligent agent's social ability test report.

[0023] The present invention provides a method for developing an intelligent agent social ability testing system, which is characterized by comprising:

[0024] S100, through 3D modeling and Unreal Engine, obtains static city scenes and develops dynamic functions of digital cities to build 3D digital cities;

[0025] S200, in the digital city, build memory-inductive interactive agents and construct social network models;

[0026] S300, select the embodied person, connect to the digital city, start the interpersonal trust algorithm expert model, and speed up the interaction training of the memory induction interactive agent and the test agent;

[0027] S400, synchronizes observation data to the intelligent agent to be tested and the expert model of the interpersonal trust algorithm, scores the feedback content of the intelligent agent to be tested, and outputs a test report on the social ability of the intelligent agent.

[0028] Preferably, S100 includes:

[0029] S101, by building a 3D model of a digital city in a 3D modeling tool, and designing and building a static city scene;

[0030] S102, define and develop dynamic functions of digital cities using Unreal Engine.

[0031] Preferably, S200 includes:

[0032] S201, building a memory-inductive interactive agent based on a large language model;

[0033] S202, in the digital city, construct a social network model based on the social network structure;

[0034] In a digital city, a social network model is constructed based on a social network structure, including: presetting multiple types of character information, importing the preset multiple types of character information into a development program in the form of a data table, configuring a mapping relationship between the character information and the character model through the development program, and generating embodied characters that conform to the preset multiple types of character information in a corresponding area in a digital city scene; constructing a social network model based on the social network structure; the social network structure includes: class structure, family structure, regional structure and employment structure; multiple types of character information include: gender information, age information, appearance information, occupation information, address information and social relationship information.

[0035] Preferably, S300 includes:

[0036] S301, the intelligent agent to be tested is connected to the digital city as an embodied person, and the interpersonal trust algorithm expert model is started at the same time;

[0037] S302, speeding up the interaction training speed of the intelligent agent to be tested by increasing the interaction frequency between the interaction request and the intelligent agent to be tested.

[0038] Preferably, S400 includes:

[0039] S401, the test system synchronizes embodied observation data to the intelligent agent to be tested and the expert model of the interpersonal trust algorithm at a set frequency; the embodiment performs interactive response actions according to the feedback of the intelligent agent to be tested; and the expert model of the interpersonal trust algorithm gives feedback;

[0040] S402, setting the number of interactive training cycles, taking the number of interactive training cycles as a node, ending the test after multiple rounds of interaction are completed, scoring the feedback content of the intelligent agent to be tested, and outputting the intelligent agent social ability test report.

[0041] Compared with the prior art, the present invention has at least the following beneficial effects:

[0042] The present invention provides an intelligent agent social ability test system and development method, which adopts a three-dimensional modeling dynamic function construction subsystem, obtains static city scenes and develops digital city dynamic functions through three-dimensional modeling and Unreal Engine, and builds a three-dimensional digital city; a social network interaction model subsystem, in the digital city, constructs a memory-inductive interaction intelligent agent and a social network model; an intelligent agent access acceleration training subsystem, selects embodied characters from the memory-inductive interaction intelligent agent, accesses the digital city, starts an interpersonal trust algorithm expert model at the same time, and accelerates the interaction training speed of the memory-inductive interaction intelligent agent and the intelligent agent to be tested; an embodied observation training output subsystem, synchronizes observation data to the intelligent agent to be tested and the interpersonal trust algorithm expert model, scores the feedback content of the intelligent agent to be tested, and outputs an intelligent agent social ability test report; the three-dimensional digital city is built based on various assets, simulates the planning and construction, social structure, various situational rules and interactive capabilities of people and objects in a real city, and provides rich and diverse scenes and elements to support social activities; the activity subject of the digital city is the intelligent agent; the intelligent agent performs situation recognition and interaction according to its own model capabilities through the input of basic role information and continuous observation data of the controlled virtual character Interactive decision-making, output behavioral instructions to virtual characters; among them, the general intelligent agent is a general intelligent agent developed in this test system that has memory, induction and planning capabilities based on a large language model; this intelligent agent mainly plays a role in the digital city according to the preset role information to build the social structure of the digital city and maintain the basic social activities of the digital city; the intelligent agent to be tested is the main test object of the test system and must have basic information interaction interface, natural language processing ability and behavioral decision-making ability; the virtual character in the digital city is the embodiment of the intelligent agent and the interface for the intelligent agent to observe the world and communicate; the intelligent agent observes and interacts in the digital city by controlling the virtual character; the expert model includes an expert model for social decision-making based on the interpersonal trust algorithm; the model is highly consistent with the real person test results in the social ability situation test, and provides a benchmark template for the social decision-making of the intelligent agent; the data recording module records the feedback results of the intelligent agent and the expert model on the interaction request in the data table; after multiple rounds of interaction, the feedback data of the data recording module is compared and analyzed, and the feedback content of the intelligent agent to be tested is scored based on the feedback results of the expert model, so as to output the social ability test report of the intelligent agent to be tested.

[0043] The present invention describes an intelligent agent social ability testing system and development method. Other advantages, objectives and features of the present invention will be reflected in part through the following description, and in part will be understood by technical personnel in the field through research and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0045] Figure 1 This is a diagram of an embodiment of an intelligent agent social ability testing system described in the present invention.

[0046] Figure 2 This is another embodiment diagram of an intelligent agent social ability testing system described in the present invention.

[0047] Figure 3 This is a diagram of a simulation operation embodiment of an intelligent agent social ability testing system and development method described in the present invention. DETAILED DESCRIPTION

[0048] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments, so that those skilled in the art can implement it according to the instructions; as shown in the figure, the present invention provides an intelligent agent social ability testing system, including:

[0049] The 3D modeling dynamic function construction subsystem uses 3D modeling and Unreal Engine to obtain static city scenes and develop digital city dynamic functions to build a 3D digital city;

[0050] The social network interaction model subsystem builds memory-inductive interaction agents and social network models in the digital city.

[0051] The agent accesses the accelerated training subsystem, selects an embodied character from the memory-inductive interactive agent, accesses it to the digital city, starts the interpersonal trust algorithm expert model, and speeds up the interactive training of the memory-inductive interactive agent and the agent to be tested;

[0052] The embodied observation training output subsystem synchronizes observation data to the intelligent agent to be tested and the expert model of the interpersonal trust algorithm, scores the feedback content of the intelligent agent to be tested, and outputs a test report on the social ability of the intelligent agent.

[0053] The principle and effect of the above technical solution are as follows: the present invention provides a system for testing the social ability of intelligent agents, including: a three-dimensional modeling dynamic function construction subsystem, which obtains static city scenes and develops dynamic functions of digital cities through three-dimensional modeling and Unreal Engine, and builds a three-dimensional digital city; a social network interaction model subsystem, which constructs a memory-inductive interaction intelligent agent and a social network model in the digital city; an intelligent agent access acceleration training subsystem, which selects embodied characters from the memory-inductive interaction intelligent agent, connects to the digital city, and simultaneously starts the interpersonal trust algorithm expert model, and accelerates the interaction training speed of the memory-inductive interaction intelligent agent and the intelligent agent to be tested; an embodied observation training output subsystem, which synchronizes observation data to the intelligent agent to be tested and the interpersonal trust algorithm expert model, scores the feedback content of the intelligent agent to be tested, and outputs an intelligent agent social ability test report; the three-dimensional digital city is built based on various assets, simulating the planning and construction, social structure, various situational rules and interactive capabilities of people and objects in a real city, and providing rich and diverse scenes and elements to support social activities; the main activity of the digital city is the intelligent agent; the intelligent agent conducts social activities according to its own model capabilities through the input of basic role information and continuous observation data of the controlled virtual character The general agent is a general agent developed in this test system that has the ability of memory, induction and planning based on a large language model. The agent mainly plays roles in the digital city according to the preset role information to build the social structure of the digital city and maintain the basic social activities of the digital city. The agent to be tested is the main test object of the test system and must have basic information interaction interface, natural language processing ability and behavior decision-making ability. The virtual character in the digital city is the embodiment of the agent and the interface for the agent to observe the world and communicate. The agent observes and interacts in the digital city by controlling the virtual character. The expert model includes an expert model for social decision-making based on the interpersonal trust algorithm. The model is highly consistent with the real person test results in the social ability situation test, providing a benchmark template for the social decision-making of the agent. The data recording module records the feedback results of the agent and the expert model on the interaction request in the data table. After multiple rounds of interaction, the feedback data of the data recording module is compared and analyzed, and the feedback content of the agent to be tested is scored based on the feedback results of the expert model, so as to output the social ability test report of the agent to be tested. Figure 1As shown, a three-dimensional digital city is constructed; the behavior instructions of the intelligent agent to be tested and the behavior instructions of the expert model are respectively transmitted to the expert feedback data center of the data recording module and the feedback data center of the intelligent agent to be tested; the expert feedback data center of the data recording module and the feedback data center of the intelligent agent to be tested are extracted to generate a social ability report of the intelligent agent to be tested; the behavior instructions of the intelligent agent to be tested are transmitted to the virtual character, and the role information of the virtual character and the observation data of the virtual character are fed back to the intelligent agent to be tested for interactive upgrade of the intelligent agent to be tested; the role information of the virtual character and the observation data of the virtual character are fed back to the expert model for analysis of the expert model of the virtual character; the virtual character and other multiple virtual characters perform social behavior simulation interaction, and the role information of multiple virtual characters and the observation data of virtual characters are fed back to the general intelligent agent; and the behavior instructions of the general intelligent agent are received; the behavior instructions of the intelligent agent to be tested are carried out to test and compare and analyze the social ability of the intelligent agent to be tested; as shown Figure 2 As shown, a three-dimensional digital city is built based on digital city art assets; a social structure is constructed in the digital city; the intelligent agent to be tested is connected to the digital city with a virtual character as an embodiment, and the expert model is started at the same time; the general intelligent agent increases the frequency of interaction with the intelligent agent to be tested; the test system synchronously observes data to the intelligent agent to be tested and the expert model, and records the feedback results; after multiple rounds of interaction, the feedback data is compared and analyzed, and a test report on the social ability of the intelligent agent is output.

[0054] In one embodiment, the 3D modeling dynamic function building subsystem includes:

[0055] The static city scene construction module builds a digital city 3D model in a 3D modeling tool and designs and builds static city scenes.

[0056] Digital city dynamic function module, using Unreal Engine to define and develop digital city dynamic functions.

[0057] The principle and effect of the above technical solution are as follows: 3D modeling dynamic function construction subsystem, including: static city scene construction module, by building a digital city 3D model in a 3D modeling tool, and designing and building a static city scene; digital city dynamic function module, using Unreal Engine to define and develop digital city dynamic functions; by building a digital city 3D model in a 3D modeling tool, and designing and building a static city scene, including: based on digital city art assets, by building a digital city 3D model in a 3D modeling tool, and designing and building a static city scene, including: based on digital city art assets, first create digital assets, and finally organize the corresponding static assets according to the designed digital city map. Arrangement to form a static city scene; digital assets include: digital buildings, digital streets, digital houses, multiple types of indoor and outdoor scenes, digital models of various common items, and character models and animations of all ages; use Unreal Engine to define and develop dynamic functions of digital cities, including: using Unreal Engine to customize the digital city navigation area and develop path-finding functions, develop and formulate corresponding traffic rules and navigation functions for different roads; develop and configure interactive functions of objects, such as controlling the movement and rotation of objects, controlling the switching of lights, and communicating with people through mobile phones; use character models and animations to configure various character behaviors and actions, and switch to corresponding actions according to changes in character status; at the same time, develop embodied characters to observe the surrounding environment and interact with others or objects.

[0058] In one embodiment, the social network interaction model subsystem includes:

[0059] The interactive agent model module builds a memory-inductive interactive agent based on a large language model;

[0060] Social network construction module, in the digital city, builds a social network model based on the social network structure;

[0061] In a digital city, a social network model is constructed based on a social network structure, including: presetting multiple types of character information, importing the preset multiple types of character information into a development program in the form of a data table, configuring a mapping relationship between the character information and the character model through the development program, and generating embodied characters that conform to the preset multiple types of character information in a corresponding area in a digital city scene; constructing a social network model based on the social network structure; the social network structure includes: class structure, family structure, regional structure and employment structure; multiple types of character information include: gender information, age information, appearance information, occupation information, address information and social relationship information.

[0062] The principle and effect of the above technical solution are as follows: the social network interaction model subsystem includes: an interactive agent model module, which constructs a memory-inductive interactive agent based on a large language model; a social network construction module, which constructs a social network model in a digital city according to a social network structure; in a digital city, constructing a social network model according to a social network structure includes: presetting multiple types of character information, importing the preset multiple types of character information into a development program in the form of a data table, configuring the mapping relationship between the character information and the character model through the development program, and generating embodied characters that meet the preset multiple types of character information in the corresponding area of ​​the digital city scene; constructing a social network model according to the social network structure; the social network structure includes: class structure, family structure, regional structure and employment structure; multiple types of character information include: gender information, age information, appearance information, occupation information, address information and social relations information; construct a memory-inductive interactive agent based on a large language model; when the test system is started, the memory-inductive interactive agent is connected to the system, and the agent obtains the basic information of the embodied person and the social relationship information through the interface input, combines the current input observation information with historical memory, generates decisions, and then controls the state changes of the characters in the scene through the interface output, and realizes interaction with the environment and real people; constructing a memory-inductive interactive agent based on a large language model includes: first, using a vector database to store and retrieve the original observation data, constructing the original memory data of the agent, and then periodically summarizing and analyzing these original memory data, extracting high-value memory data and constructing it into important memory data, and finally performing situational reasoning and interactive decision-making based on the timeliness, importance and relevance of the memory; the agent has memory-inductive interaction capabilities and multiple identity recognition functions based on interpersonal relationships.

[0063] In one embodiment, the agent accesses the accelerated training subsystem, including:

[0064] The embodied access model startup module, in which the intelligent agent to be tested is connected to the digital city as an embodied person, and the interpersonal trust algorithm expert model is started at the same time;

[0065] The interaction frequency acceleration training module speeds up the interaction training of the intelligent agent to be tested by increasing the interaction frequency of interaction requests and the intelligent agent to be tested.

[0066] The principle and effect of the above technical solution are as follows: the intelligent agent access acceleration training subsystem includes: the embodied access model startup module, the intelligent agent to be tested is connected to the digital city as an embodied person, and the interpersonal trust algorithm expert model is started at the same time; the interaction frequency acceleration training module speeds up the interaction training speed of the intelligent agent to be tested by increasing the interaction frequency of interaction requests with the intelligent agent to be tested; the intelligent agent to be tested is connected to the digital city as an embodied person, and the interpersonal trust algorithm expert model is started at the same time, including: when the test starts, the intelligent agent to be tested selects any embodied person as an embodied person to be connected to the environment, and the test system inputs the embodied person's characteristic information to the intelligent agent to be tested through the interface; the embodied person's characteristic information includes: the embodied person's basic information, the embodied person's ability information and the social relationship information; the intelligent agent to be tested performs data processing according to the embodied person's ability information; at the same time, the test system will The characteristic information is input into the interpersonal trust algorithm expert model; the interpersonal trust algorithm expert model obtains the trust level through the interpersonal trust algorithm, makes decisions based on the trust level combined with the interaction request and the context, and outputs feedback data; in the subsequent testing process, the test system will send the data sent to the intelligent agent to be tested to the interpersonal trust algorithm expert model synchronously; by increasing the interaction frequency of the interaction request and the intelligent agent to be tested, the interaction training speed of the intelligent agent to be tested is accelerated, including: testing the social ability of the intelligent agent to be tested, significantly improving the interaction frequency with the intelligent agent to be tested, that is, issuing more interaction requests to the intelligent agent to be tested per unit time; issuing multiple types of interaction requests to the intelligent agent to be tested according to the interaction influencing factors; the request forms include: whether the question form, the selection question form and the description question form; the interaction influencing factors include: environmental changes, related memories, preset conditions and external guiding factors.

[0067] In one embodiment, the embodied observation training output subsystem includes:

[0068] Synchronous embodied observation module: the test system synchronizes embodied observation data to the tested intelligent agent and the interpersonal trust algorithm expert model at a set frequency; the embodiment performs interactive response actions based on the feedback from the tested intelligent agent; and the interpersonal trust algorithm expert model gives feedback;

[0069] The training cycle scoring output module sets the number of interactive training cycles, takes the number of interactive training cycles as a node, ends the test after multiple rounds of interaction are completed, scores the feedback content of the tested intelligent agent, and outputs the intelligent agent's social ability test report.

[0070] The principles and effects of the above technical solution are: the embodied observation training output subsystem includes: a synchronous embodied observation module, in which the test system synchronizes embodied observation data to the intelligent agent to be tested and the expert model of interpersonal trust algorithm at a set frequency; the embodiment performs interactive response actions based on the feedback from the intelligent agent to be tested; the expert model of interpersonal trust algorithm gives feedback; a training cycle scoring output module, which sets the number of interactive training cycles, takes the number of interactive training cycles as a node, ends the test after multiple rounds of interaction are completed, and scores the feedback content of the intelligent agent to be tested, and outputs a test report on the social ability of the intelligent agent; the test system synchronizes embodied observation data to the intelligent agent to be tested and the expert model of interpersonal trust algorithm at a set frequency; the embodiment performs interactive response actions based on the feedback from the intelligent agent to be tested; the expert model of interpersonal trust algorithm gives feedback including: the test system synchronizes embodied observation data to the intelligent agent to be tested and the expert model of interpersonal trust algorithm at a set frequency The model synchronizes embodied observation data; embodied observation data includes environmental observation data and agent request content; when an interaction request is triggered, the agent to be tested gives feedback based on its own capabilities, and the embodiment performs an interactive response action based on this; the interpersonal trust algorithm expert model gives feedback based on the trust judgment of the requester and the current situation; the test system records the feedback results of the agent and the interpersonal trust algorithm expert model in the data table; the setting frequency includes: 6 times / minute or other adaptive frequencies; trust algorithm decision model module, the interpersonal trust algorithm expert model is a model for social decision-making based on the interpersonal trust algorithm. The model takes trust as the core element for social decision-making. It analyzes the relationship between characters and interaction records and combines its own values ​​to judge the degree of trust in the interaction requester, and then considers the rationality and difficulty of the request content to itself, so as to make social decisions;

[0071] The specific calculation formula of trust is as follows:

[0072]

[0073] in, Indicates the trust score. When the score is positive, the response to the interaction request is yes. When the score is negative, the response to the interaction request is no. Indicates the intimacy value with the requester, ranging from 0 to 1 (0 is completely unfamiliar, 1 is extremely close); Indicates the interaction value with the requester in the recent period, ranging from 0 to 1 (0 means no contact, 1 means close contact); Indicates the character base of the virtual character, ranging from 0 to 1 (0 is completely indifferent, 1 is extremely enthusiastic); Indicates the objective difficulty of the interaction request, ranging from 0 to 1 (0 means no difficulty, 1 means impossible to achieve); Indicates the situational rationality of the interaction request, ranging from 0 to 1 (0 is completely reasonable and 1 is completely unreasonable).

[0074] The present invention provides a method for developing an intelligent agent social ability testing system, which is characterized by comprising:

[0075] S100, through 3D modeling and Unreal Engine, obtains static city scenes and develops dynamic functions of digital cities to build 3D digital cities;

[0076] S200, in the digital city, build memory-inductive interactive agents and build social network models;

[0077] S300, select the embodied person, connect to the digital city, start the interpersonal trust algorithm expert model, and speed up the interaction training of the memory induction interactive agent and the test agent;

[0078] S400, synchronizes observation data to the intelligent agent to be tested and the expert model of the interpersonal trust algorithm, scores the feedback content of the intelligent agent to be tested, and outputs a test report on the social ability of the intelligent agent.

[0079] The principle and effect of the above technical solution are as follows: the present invention provides a method for developing a social ability test system for an intelligent agent, comprising: obtaining static city scenes and developing dynamic functions of a digital city through three-dimensional modeling and Unreal Engine, and building a three-dimensional digital city; in the digital city, constructing a memory-inductive interactive intelligent agent and a social network model; selecting an embodied person, accessing the digital city, and simultaneously starting an expert model of an interpersonal trust algorithm, and accelerating the interactive training speed of the memory-inductive interactive intelligent agent and the intelligent agent to be tested; synchronizing observation data to the intelligent agent to be tested and the expert model of the interpersonal trust algorithm, and scoring the feedback content of the intelligent agent to be tested, and outputting an intelligent agent social ability test report; the three-dimensional digital city is built based on various assets, simulating the planning and construction, social structure, various situational rules and interactive capabilities of people and objects in a real city, and providing rich and diverse scenarios and elements to support social activities; the main activity of the digital city is an intelligent agent; the intelligent agent, through the input of the basic role information and continuous observation data of the controlled virtual character, performs situation recognition and interactive decision-making according to its own model capabilities, and outputs behavioral instructions to the virtual character; wherein, the general intelligent The agent is a general agent developed in this test system that has the ability of memory, induction and planning based on a large language model; the agent mainly plays roles in the digital city according to the preset role information to build the social structure of the digital city and maintain the basic social activities of the digital city; the agent to be tested is the main test object of the test system and must have basic information interaction interface, natural language processing ability and behavioral decision-making ability; the virtual characters in the digital city are the embodiment of the agent and the interface for the agent to observe the world and communicate; the agent observes and interacts in the digital city by controlling the virtual characters; the expert model includes an expert model for social decision-making based on the interpersonal trust algorithm; the model is highly consistent with the real person test results in the social ability situation test, and provides a benchmark template for the social decision-making of the agent; the data recording module records the feedback results of the agent and the expert model on the interaction request into the data table; after multiple rounds of interaction, the feedback data of the data recording module is compared and analyzed, and the feedback content of the agent to be tested is scored based on the feedback results of the expert model, so as to output the social ability test report of the agent to be tested; Figure 1As shown, a three-dimensional digital city is constructed; the behavior instructions of the intelligent agent to be tested and the behavior instructions of the expert model are respectively transmitted to the expert feedback data center of the data recording module and the feedback data center of the intelligent agent to be tested; the expert feedback data center of the data recording module and the feedback data center of the intelligent agent to be tested are extracted to generate a social ability report of the intelligent agent to be tested; the behavior instructions of the intelligent agent to be tested are transmitted to the virtual character, and the role information of the virtual character and the observation data of the virtual character are fed back to the intelligent agent to be tested for interactive upgrade of the intelligent agent to be tested; the role information of the virtual character and the observation data of the virtual character are fed back to the expert model for analysis of the expert model of the virtual character; the virtual character and other multiple virtual characters perform social behavior simulation interaction, and the role information of multiple virtual characters and the observation data of virtual characters are fed back to the general intelligent agent; and the behavior instructions of the general intelligent agent are received; the behavior instructions of the intelligent agent to be tested are carried out to test and compare and analyze the social ability of the intelligent agent to be tested; as shown Figure 2 As shown, a three-dimensional digital city is built based on digital city art assets; a social structure is constructed in the digital city; the intelligent agent to be tested is connected to the digital city with a virtual character as an embodiment, and the expert model is started at the same time; the general intelligent agent increases the frequency of interaction with the intelligent agent to be tested; the test system synchronously observes data to the intelligent agent to be tested and the expert model, and records the feedback results; after multiple rounds of interaction, the feedback data is compared and analyzed, and a test report on the social ability of the intelligent agent is output.

[0080] In one embodiment, S100 includes:

[0081] S101, by building a 3D model of a digital city in a 3D modeling tool, and designing and building a static city scene;

[0082] S102, define and develop dynamic functions of digital cities using Unreal Engine.

[0083] The principle and effect of the above technical solution are: by constructing a three-dimensional model of a digital city in a three-dimensional modeling tool, and designing and building a static city scene; using the Unreal Engine to define and develop dynamic functions of a digital city; by constructing a three-dimensional model of a digital city in a three-dimensional modeling tool, and designing and building a static city scene, including: based on digital city art assets, by constructing a three-dimensional model of a digital city in a three-dimensional modeling tool, and designing and building a static city scene, including: based on digital city art assets, first create digital assets, and finally according to the designed digital city map, organize and arrange the corresponding static assets to form a static city scene; digital assets include: Digital buildings, digital streets, digital houses, multiple types of indoor and outdoor scenes, digital models of various common objects, and character models and animations of all ages; using Unreal Engine to define and develop dynamic functions of digital cities, including: using Unreal Engine to customize the digital city navigation area and develop path-finding functions, develop and formulate corresponding traffic rules and navigation functions for different roads; develop and configure interactive functions of objects, such as controlling the movement and rotation of objects, controlling the switching of lights, and communicating with people through mobile phones; using character models and animations to configure various character behaviors and actions, and switching to corresponding actions according to changes in the character's state; and developing interfaces for embodied characters to observe the surrounding environment and interact with others or objects.

[0084] In one embodiment, S200 includes:

[0085] S201, building a memory-inductive interactive agent based on a large language model;

[0086] S202, in the digital city, construct a social network model based on the social network structure;

[0087] In a digital city, a social network model is constructed based on a social network structure, including: presetting multiple types of character information, importing the preset multiple types of character information into a development program in the form of a data table, configuring a mapping relationship between the character information and the character model through the development program, and generating embodied characters that conform to the preset multiple types of character information in a corresponding area in a digital city scene; constructing a social network model based on the social network structure; the social network structure includes: class structure, family structure, regional structure and employment structure; multiple types of character information include: gender information, age information, appearance information, occupation information, address information and social relationship information.

[0088] The principle and effect of the above technical solution are: based on the large language model, a memory-inductive interactive intelligent body is constructed; in the digital city, a social network model is constructed according to the social network structure; in the digital city, according to the social network structure, the social network model is constructed including: presetting multiple types of character information, importing the preset multiple types of character information into the development program in the form of a data table, configuring the mapping relationship between the character information and the character model through the development program, and generating embodied characters that meet the preset multiple types of character information in the corresponding area of ​​the digital city scene; constructing a social network model according to the social network structure; the social network structure includes: class structure, family structure, regional structure and employment structure; multiple types of character information include: gender information, age information, appearance information, occupation information, address information and social relationship information; constructing a memory-inductive interactive intelligent body based on the large language model Interactive agent; when the test system is started, the memory-inductive interactive agent is connected to the system. The agent obtains the basic information of the embodied character and social relationship information through the interface input, combines the current input observation information with historical memory, generates decisions, and then controls the status changes of the characters in the scene through the interface output, and realizes interaction with the environment and real people; Building a memory-inductive interactive agent based on a large language model includes: first, using a vector database to store and retrieve the original observation data, constructing the original memory data of the agent, and then periodically summarizing and analyzing these original memory data, extracting high-value memory data and constructing it into important memory data, and finally performing situational reasoning and interactive decision-making based on the timeliness, importance and relevance of the memory; The agent has memory-inductive interaction capabilities and multiple identity recognition functions based on interpersonal relationships.

[0089] In one embodiment, S300 includes:

[0090] S301, the intelligent agent to be tested is connected to the digital city as an embodied person, and the interpersonal trust algorithm expert model is started at the same time;

[0091] S302, speeding up the interaction training speed of the intelligent agent to be tested by increasing the interaction frequency between the interaction request and the intelligent agent to be tested.

[0092] The principle and effect of the above technical solution are as follows: the intelligent agent to be tested is connected to the digital city as an embodied person, and the expert model of interpersonal trust algorithm is started at the same time; by increasing the interaction frequency of interaction requests and the intelligent agent to be tested, the interaction training speed of the intelligent agent to be tested is accelerated; the intelligent agent to be tested is connected to the digital city as an embodied person, and the expert model of interpersonal trust algorithm is started at the same time, including: when the test starts, the intelligent agent to be tested selects any embodied person as an embodied person to be connected to the environment, and the test system inputs the embodied person's characteristic information to the intelligent agent to be tested through the interface; the embodied person's characteristic information includes: the embodied person's basic information, the embodied person's ability information and the social relationship information; the intelligent agent to be tested performs data processing according to the embodied person's ability information; at the same time, the test system inputs the embodied person's characteristic information into the expert model of interpersonal trust algorithm; The interpersonal trust algorithm expert model obtains the trust level through the interpersonal trust algorithm, makes decisions based on the trust level combined with the interaction request and the context, and outputs feedback data; in the subsequent testing process, the test system will send the data sent to the tested intelligent agent to the interpersonal trust algorithm expert model synchronously; by increasing the interaction frequency of the interaction request and the tested intelligent agent, the interaction training speed of the tested intelligent agent is accelerated, including: testing the social ability of the tested intelligent agent, significantly improving the interaction frequency with the tested intelligent agent, that is, issuing more interaction requests to the tested intelligent agent per unit time; issuing multiple types of interaction requests to the tested intelligent agent according to the interaction influencing factors; the request forms include: whether the question form, the selection question form and the description question form; the interaction influencing factors include: environmental changes, related memories, preset conditions and external guidance factors.

[0093] In one embodiment, S400 includes:

[0094] S401, the test system synchronizes embodied observation data to the intelligent agent to be tested and the expert model of the interpersonal trust algorithm at a set frequency; the embodiment performs interactive response actions according to the feedback of the intelligent agent to be tested; and the expert model of the interpersonal trust algorithm gives feedback;

[0095] S402, setting the number of interactive training cycles, taking the number of interactive training cycles as a node, ending the test after multiple rounds of interaction are completed, scoring the feedback content of the intelligent agent to be tested, and outputting the intelligent agent social ability test report.

[0096] The principle and effect of the above technical solution are: the test system synchronizes embodied observation data to the intelligent agent to be tested and the expert model of interpersonal trust algorithm at a set frequency; the embodiment performs interactive response actions according to the feedback of the intelligent agent to be tested; the expert model of interpersonal trust algorithm gives feedback; the number of interactive training cycles is set, and the number of interactive training cycles is used as a node. After multiple rounds of interaction are completed, the test ends, and the feedback content of the intelligent agent to be tested is scored, and a test report on the social ability of the intelligent agent is output; the test system synchronizes embodied observation data to the intelligent agent to be tested and the expert model of interpersonal trust algorithm at a set frequency; the embodiment performs interactive response actions according to the feedback of the intelligent agent to be tested; the expert model of interpersonal trust algorithm gives feedback, including: the test system synchronizes embodied observation data to the intelligent agent to be tested and the expert model of interpersonal trust algorithm at a set frequency; the embodied observation data includes loop The test system collects the environment observation data and agent request content; when the interaction request is triggered, the agent to be tested gives feedback based on its own capabilities, and the embodiment performs the interactive response action according to the feedback of the agent to be tested; the interpersonal trust algorithm expert model gives feedback based on the trust judgment of the requester and the current situation; the test system records the feedback results of the agent and the interpersonal trust algorithm expert model in the data table; the setting frequency includes: 6 times / minute or other adaptation frequency; the trust algorithm decision model module, the interpersonal trust algorithm expert model is a model for social decision-making based on the interpersonal trust algorithm. The model takes trust as the core element for social decision-making. It analyzes the relationship between people and the interaction records and judges the trustworthiness of the interaction requester in combination with its own values, and then considers the rationality and difficulty of the request content to itself, so as to make social decisions;

[0097] The specific calculation formula of trust is as follows:

[0098]

[0099] in, Indicates the trust score. When the score is positive, the response to the interaction request is yes. When the score is negative, the response to the interaction request is no. Indicates the intimacy value with the requester, ranging from 0 to 1 (0 is completely unfamiliar, 1 is extremely close); Indicates the interaction value with the requester in the recent period, ranging from 0 to 1 (0 means no contact, 1 means close contact); Indicates the character base of the virtual character, ranging from 0 to 1 (0 is completely indifferent, 1 is extremely enthusiastic); Indicates the objective difficulty of the interaction request, ranging from 0 to 1 (0 means no difficulty, 1 means impossible to achieve); Indicates the situational rationality of the interaction request, ranging from 0 to 1 (0 is completely reasonable and 1 is completely unreasonable).

[0100] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and the implementation methods. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and the illustrations shown and described herein.

Claims

1. A system for testing the social ability of an intelligent agent, characterized in that: include: The 3D modeling dynamic function construction subsystem uses 3D modeling and Unreal Engine to obtain static city scenes and develop digital city dynamic functions to build a 3D digital city; The social network interaction model subsystem builds memory-inductive interaction agents and social network models in the digital city. The intelligent agent is connected to the accelerated training subsystem, and a virtual character is selected as the intelligent agent to be tested and connected to the digital city; the virtual character in the digital city is the embodiment of the memory-inductive interactive intelligent agent; at the same time, the interpersonal trust algorithm expert model is started, and the interactive training speed of the memory-inductive interactive intelligent agent and the intelligent agent to be tested is accelerated; The embodied observation training output subsystem synchronously transmits embodied observation data to the tested intelligent agent and the interpersonal trust algorithm expert model. The embodied observation data includes environmental observation data and interaction request content; records the feedback data of the tested intelligent agent, scores the feedback content of the tested intelligent agent based on the feedback results of the expert model, and outputs a test report on the social ability of the tested intelligent agent; The social network interaction model subsystem includes: building a memory-inductive interaction agent based on a large language model; The memory-inductive interactive agent plays roles in the digital city according to preset role information, constructs the social network structure of the digital city, maintains the basic social activities of the digital city, and obtains a social network model.

2. A method for developing an intelligent agent social ability testing system, characterized in that: include: S100, through 3D modeling and Unreal Engine, obtains static city scenes and develops dynamic functions of digital cities to build 3D digital cities; S200, in the digital city, build memory-inductive interactive agents and build social network models; S300, selecting a virtual person as the subject to be tested and connecting it to the digital city; the virtual person in the digital city is the embodiment of the memory-inductive interaction agent; at the same time, starting the interpersonal trust algorithm expert model, and accelerating the interaction training speed between the memory-inductive interaction agent and the subject to be tested; S400, synchronously transmitting embodied observation data to the intelligent agent to be tested and the expert model of the interpersonal trust algorithm, the embodied observation data including environmental observation data and interaction request content; recording the feedback data of the intelligent agent to be tested, taking the feedback results of the expert model as the standard, and scoring the feedback content of the intelligent agent to be tested, and outputting a social ability test report of the intelligent agent to be tested; S200 includes: building a memory-inductive interactive agent based on a large language model; The memory-inductive interactive agent plays roles in the digital city according to preset role information, constructs the social network structure of the digital city, maintains the basic social activities of the digital city, and obtains a social network model.

Images