A method, device, equipment and storage medium for generating parent-child interaction decision information

Abstract

The present invention discloses a method, device, equipment and storage medium for generating parent-child interaction decision information. The method comprises: obtaining interaction context data, child demand data and parent education style data; inputting the interaction context data, child demand data and parent education style data into an intelligent agent role allocation model to obtain multiple intelligent agents and the weight of each intelligent agent; inputting the multiple intelligent agents and the weight of each intelligent agent into an intelligent agent collaborative decision model to obtain interaction decision information; displaying the interaction decision information to parents and children, and obtaining interaction feedback information after parents and children interact according to the interaction decision; inputting the interaction feedback information into a cyclic iterative optimization algorithm to obtain updated interaction decision information. The present invention realizes the continuous optimization of parent-child interaction suggestions through the collaborative work of multiple role intelligent agents. By introducing the intelligent agent role allocation model, the roles and functions of the intelligent agents can be dynamically adjusted according to different interaction needs.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method, device, equipment and storage medium for generating parent-child interaction decision-making information. Background Art

[0002] With the rapid development of artificial intelligence technology, intelligent assistance systems are being used more and more widely in the field of education, showing significant potential especially in parent-child interaction scenarios.

[0003] However, existing assistance systems generally lack an iterative mechanism, limiting their long-term effectiveness and adaptability. Most current parent-child interaction assistance systems employ a one-way recommendation model, generating interaction suggestions based on preset rules or static models and then directly outputting them. These systems are unable to dynamically adjust strategies through continuous data feedback and algorithmic optimization. This one-way decision chain makes it difficult for the system to respond to changes in user behavior, environmental dynamics, or individual needs. For example, it is unable to adaptively update parameters based on the real-time progress of a child's development or adjustments to family interaction patterns.

[0004] Furthermore, system parameters often rely on offline training or manual intervention, lacking online learning capabilities, causing recommendation results to gradually deviate from actual needs. This static design not only weakens the system's exploration capabilities (such as trying potentially better interaction solutions), but also causes recommendation strategies to become trapped in local optima, making it difficult to break through the limitations of the initial settings. Summary of the Invention

[0005] The purpose of the present invention is to provide a method, device, equipment and storage medium for generating parent-child interaction decision-making information, aiming to solve the common problems of existing auxiliary systems such as the lack of a loop iteration mechanism.

[0006] In a first aspect, an embodiment of the present invention provides a method for generating parent-child interaction decision information, comprising:

[0007] Obtain interaction context data, children's needs data, and parents' educational style data;

[0008] Inputting the interaction context data, child demand data, and parental education style data into an agent role allocation model to obtain multiple agents and the weight of each agent;

[0009] Input multiple agents and the weights of each agent into the agent collaborative decision-making model to obtain interactive decision information;

[0010] Presenting the interactive decision information to the parent and the child, and obtaining interactive feedback information from the parent and the child after the interaction according to the interactive decision;

[0011] The interactive feedback information is input into a cyclic iterative optimization algorithm to obtain updated interactive decision information.

[0012] In a second aspect, an embodiment of the present invention provides a parent-child interaction decision information generating device, comprising:

[0013] A data acquisition unit, used to acquire interaction context data, children's needs data, and parents' education style data;

[0014] An agent acquisition unit, configured to input the interaction context data, child demand data, and parental education style data into an agent role allocation model to obtain multiple agents and the weight of each agent;

[0015] A decision acquisition unit is used to input multiple agents and the weights of each agent into the agent collaborative decision model to obtain interactive decision information;

[0016] A display unit, configured to display the interactive decision information to the parent and the child, and obtain interactive feedback information from the parent and the child after the interaction according to the interactive decision;

[0017] The updating unit is used to input the interactive feedback information into the cyclic iterative optimization algorithm to obtain updated interactive decision information.

[0018] In a third aspect, an embodiment of the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for generating parent-child interaction decision information described in the first aspect is implemented.

[0019] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for generating parent-child interaction decision information described in the first aspect above is implemented.

[0020] The present invention discloses a method, device, equipment, and storage medium for generating parent-child interaction decision information. The method comprises: obtaining interaction context data, child demand data, and parental education style data; inputting the interaction context data, child demand data, and parental education style data into an agent role allocation model to obtain multiple agents and the weights of each agent; inputting the multiple agents and the weights of each agent into an agent collaborative decision model to obtain interaction decision information; displaying the interaction decision information to parents and children, and obtaining interaction feedback information after the parents and children interact according to the interaction decision; and inputting the interaction feedback information into a cyclic iterative optimization algorithm to obtain updated interaction decision information. The present invention achieves continuous optimization of parent-child interaction suggestions through the collaborative work of multiple role agents. By introducing an agent role allocation model, the roles and functions of the agents can be dynamically adjusted according to different interaction needs, thereby providing more accurate support for parent-child interaction. The embodiments of the present invention also provide a parent-child interaction decision information generation device, a computer-readable storage medium, and a computer device, which have the above-mentioned beneficial effects and are not further described here. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0022] Figure 1 A flowchart of a method for generating decision-making information for parent-child interaction;

[0023] Figure 2 A schematic block diagram of a device for generating parent-child interaction decision information. DETAILED DESCRIPTION

[0024] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0025] It will be understood that when used in this specification and the appended claims, the terms “comprises” and “comprising” indicate the presence of described features, integers, steps, operations, elements and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or collections thereof.

[0026] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the present invention. As used in the specification and appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0027] It should be further understood that the term "and / or" used in the present description and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0028] See also Figure 1 This embodiment provides a method for generating parent-child interaction decision information, including:

[0029] S101: Acquire interaction context data, children’s needs data, and parents’ education style data;

[0030] The collection of interaction context data relies on a multimodal sensor network, including visual cameras deployed in the home environment (to capture visual information of parent-child interactions), voice acquisition modules (to record conversation content and tonal characteristics), and wearable devices (such as children's wristbands or parents' smartwatches, which monitor physiological indicators such as heart rate and movement frequency). After aligning the multi-source data using timestamp synchronization technology, the Transformer encoder extracts scene features (such as spatial layout, interaction duration, and body language). This is then combined with an environmental state recognition model (based on the YOLOv5 architecture) to annotate key interaction events (such as game start, emotional conflict, etc.), forming a structured description vector of the interaction context.

[0031] Secondly, data on children's needs is collected through a combination of behavior recognition models and sentiment analysis algorithms. The system first analyzes children's body movements (e.g., reaching out, shaking their head) using a 3D pose estimation model (such as OpenPose). This is then combined with a facial expression recognition model (such as Facenet+ResNet) to determine emotional states (e.g., joy, anxiety). Furthermore, an attention mechanism is used to model historical behavior sequences, and an LSTM network is used to predict children's potential needs (e.g., desire for interaction, need for rest). For example, if a child frequently points to a toy but receives no response, the system identifies this as an "exploration need" and encodes it as a need feature vector.

[0032] Finally, a hybrid data fusion approach was used to collect data on parenting styles. First, a questionnaire interface was used to collect parents' subjective evaluations (e.g., educational goals and patience). Natural language processing models (e.g., BERT) were then used to perform semantic analysis on parents' voice commands, extracting characteristics of parenting styles (e.g., directive versus encouraging). Second, based on behavioral log analysis, clustering algorithms (e.g., K-means) were used to categorize parents' interaction patterns (e.g., high involvement versus low intervention), and their behavioral preference weights were calculated. Finally, a multi-source feature alignment algorithm was used to fuse the subjective evaluations with the objective behavioral data to generate a parenting style feature vector.

[0033] After preprocessing, the above three types of data are input into the core module of the system (such as the cyclic iterative optimization algorithm) and the state-action feature vector is generated through the joint feature encoder. , serves as the input to the action-value function, driving subsequent strategy optimization. This embodiment significantly improves the system's ability to understand complex parent-child interaction scenarios through multi-dimensional data collection and feature modeling, laying the data foundation for the generation of dynamic recommendation strategies.

[0034] S102: Inputting the interaction scenario data, child demand data, and parental education style data into an agent role allocation model to obtain multiple agents and the weight of each agent;

[0035] Specifically, the interaction context data, children's needs data, and parents' education style data are input into the agent role allocation model to obtain multiple agents and the weights of each agent, including:

[0036] Extract features from interaction context data, child demand data, and parent education style data to generate interaction context features, child demand features, and parent education style features respectively;

[0037] The interaction context features, children's demand features and parents' educational style features are integrated into a comprehensive feature vector through the feature fusion function;

[0038] Calculate the fitness score of each role in the predefined role set based on the comprehensive feature vector;

[0039] Sort the characters according to their compatibility score and select compatible character combinations;

[0040] Normalize the fitness scores of the selected roles and generate the weights of each role;

[0041] According to the weight of each role and through the agent selection function, a corresponding agent is assigned to each role to obtain multiple agents and the weight of each agent.

[0042] This embodiment achieves accurate modeling of parent-child interaction scenarios and efficient scheduling of intelligent agent resources through feature extraction, fusion and dynamic role allocation, significantly improving the system's adaptability and collaborative efficiency.

[0043] In this embodiment, the agent role allocation model is a key component of the agent-driven parent-child interaction loop system. It is used to dynamically assign the most appropriate agent role combination based on the interaction context, child needs, and parenting style. This model uses a multi-dimensional matching algorithm to ensure that the system can call upon the most appropriate agent role combination in each parent-child interaction scenario, providing precise interaction support. This mathematical representation enables the system to flexibly adjust the agent role positioning and function combination based on the unique needs and context of each family.

[0044] Furthermore, the mathematical expression of the agent role allocation model is as follows:

[0045] AgentRole=f(InteractionContext, ChildNeeds, ParentStyle);

[0046] Among them, AgentRole∈R is the result of the agent role assignment, which represents the agent role combination selected by the system for the current interaction scenario; R represents the space of all possible role assignment results;

[0047] The role assignment result can be represented as a structured data:

[0048] AgentRole={(a1, r1, w1), (a2, r2, w2),...(ai, ri, wi)..., (an, rn, wn)};

[0049] Among them, ai∈A represents the i-th agent, which comes from the agent pool A; ri∈Rsingle represents the role assigned to agent ai, which comes from the single role set Rsingle; wi∈[0,1] represents the weight of agent ai in the current role assignment, satisfying .

[0050] Furthermore, InteractionContext∈C is the interaction context, which represents the environment and background information of the current parent-child interaction; C represents the space of all possible interaction contexts;

[0051] The interaction context can be represented as a structured data:

[0052] InteractionContext=(tContext, lContext, aContext, gContext, hContext);

[0053] Among them, t context represents the interaction type, such as reading, gaming, conversation, creation, etc.; l context represents the interaction location, such as home, school, outdoors, virtual space, etc.; a context represents the interaction activity, such as specific books, games, topics, etc.; g context represents the interaction goal, such as knowledge learning, ability development, emotional communication, etc.; h context represents the historical interaction record, which contains summary information of past related interactions.

[0054] Furthermore, ChildNeeds∈N is the child needs, which represents the current developmental, learning, and emotional needs of children; N represents the space of all possible child needs;

[0055] Children's needs can be represented as a structured data:

[0056] ChildNeeds=(dneed, ineed, eneed, sneed, cneed);

[0057] Among them, d needs represent developmental needs, which are related to the children's current development stage; i needs represent interest needs, which are related to children's interests and hobbies; e needs represent emotional needs, which are related to children's emotional state; s needs represent social needs, which are related to the development of children's social skills; c needs represent cognitive needs, which are related to the development of children's cognitive abilities.

[0058] Furthermore, ParentStyle∈P is the parenting style, which represents the parent’s preferences and methods in education and interaction; P represents the space of all possible parenting styles;

[0059] Parenting style can be represented as a structured data:

[0060] ParentStyle=(a style, r style, c style, s style, p style);

[0061] Among them, astyle∈[0,1] is authority, which indicates the degree of authority of parents in education; rstyle∈[0,1] is responsiveness, which indicates the degree of response of parents to children's needs; cstyle∈[0,1] is control, which indicates the degree of control of children's behavior by parents; sstyle∈[0,1] is support, which indicates the degree of support and encouragement of children by parents; pstyle∈Ppref is educational preference, which indicates the preference of parents on educational methods and content.

[0062] f(C×N×P→R) represents the role assignment function, which maps the interaction context, child needs, and parenting style to the resulting agent role assignment. This role assignment function can be implemented using a variety of methods, such as rule matching, similarity calculation, or machine learning models. In practical systems, the role assignment function typically consists of multiple sub-functions and processing steps.

[0063] In this embodiment, multiple agent role types are supported. Each role has a specific functional positioning and applicable scenarios. The following are some examples of role types:

[0064] 1. Educational guide role (reducer):

[0065] Functional positioning: Provide scientific teaching methods and knowledge content;

[0066] Applicable scenarios: knowledge learning, skill cultivation, cognitive development;

[0067] Eigenvectors: ;

[0068] Matching function: ;

[0069] 2. Emotional support role (remotional):

[0070] Functional positioning: focus on emotional needs, provide emotional support and adjustment;

[0071] Applicable scenarios: emotional management, strengthening parent-child relationships, and building a sense of security;

[0072] Eigenvectors: ;

[0073] Matching function: ;

[0074] 3. Game guide role (rplay):

[0075] Functional positioning: Design and guide interesting and educational game activities;

[0076] Applicable scenarios: game interaction, creative activities, exploratory learning;

[0077] Eigenvectors: ;

[0078] Matching function: ;

[0079] 4. Development Consultant Role (rdevelopment):

[0080] Functional positioning: Focus on children's development milestones and critical periods;

[0081] Applicable scenarios: development assessment, capacity building, and critical period intervention;

[0082] Eigenvectors: ;

[0083] Matching function: ;

[0084] 5. Social Coach Role (rsocial):

[0085] Functional positioning: Cultivate social skills and interpersonal interaction ability;

[0086] Applicable scenarios: social learning, conflict resolution, and cooperative activities;

[0087] Eigenvectors: ;

[0088] Matching function: ;

[0089] 6. Creative Inspirer Role (rcreative):

[0090] Functional positioning: stimulate creativity and imagination;

[0091] Applicable scenarios: artistic creation, story creation, innovative thinking;

[0092] Eigenvectors: ;

[0093] Matching function: ;

[0094] 7. Parent coach role (rparent):

[0095] Functional positioning: guide parents to improve their educational methods and interaction skills;

[0096] Applicable scenarios: parent training, educational method guidance, and parent-child interaction optimization;

[0097] Eigenvectors: ;

[0098] Matching function: ;

[0099] 8. Safety Guardian Role (rsafety):

[0100] Functional positioning: ensuring the safety and appropriateness of interactive content and methods;

[0101] Applicable scenarios: content screening, risk assessment, and safety education;

[0102] Eigenvectors: ;

[0103] Matching function: ;

[0104] In some embodiments, interaction context features may be extracted by:

[0105] ;

[0106] Among them, Encodercontext represents the context encoder, which can be a rule-based feature extractor or a neural network encoder; Represents the interaction context feature vector with dimension dc.

[0107] The interaction context feature vector can contain the following dimensions:

[0108] Interaction type feature: ftype=OneHot(t)orEmbedding(t);

[0109] Interaction location feature: flocation=OneHot(l)orEmbedding(l);

[0110] Interactive activity feature: factivity=Embedding(a);

[0111] Interactive target feature: fgoal=MultiHot(g)orEmbedding(g);

[0112] History interaction feature: fhistory=Summarizer(h).

[0113] In some embodiments, the child demand features may be extracted by:

[0114] ;

[0115] Among them, Encoderneeds represents the demand encoder, which can be a rule-based feature extractor or a neural network encoder; Represents the children's demand feature vector, with dimension dn.

[0116] The child demand feature vector can contain the following dimensions:

[0117] Development demand feature: fdevelopment=Encoderdev(d);

[0118] Interest demand feature: finterest=Encoderint(i);

[0119] Emotional demand characteristics: femotion=Encoderemo(e);

[0120] Social demand characteristics: fsocial=Encodersoc(s);

[0121] Cognitive demand characteristics: fcognitive=Encodercog(c);

[0122] In some embodiments, the parenting style features may be extracted by:

[0123] ;

[0124] Among them, Encoderstyle represents the style encoder, which can be a rule-based feature extractor or a neural network encoder; Represents the parental education style vector with dimension ds.

[0125] The parenting style vector can directly use the parameters of the parenting style:

[0126] .

[0127] Furthermore, the interaction context characteristics, children's needs characteristics, and parents' educational style characteristics are integrated into a unified representation:

[0128] ;

[0129] Among them, Fusion represents the feature fusion function, which can be a simple splicing, weighted sum or a more complex fusion network. Represents the fused feature vector, with dimension df.

[0130] Feature fusion can be performed using the following methods:

[0131] Simple splicing: ;

[0132] Weighted sum: ;

[0133] Attention Mechanism: ;

[0134] Multilayer Perceptron: .

[0135] The following is an example of the application of the agent role allocation model in a parent-child interaction scenario:

[0136] 1. Read the interactive scene:

[0137] Interaction Context: Interaction Type: Reading; Interaction Location: Family; Interaction Activity: Winnie the Pooh Storybook; Interaction Objectives: Language Development, Emotional Communication; Historical Interaction: Multiple successful parent-child reading experiences in the past;

[0138] Children's needs: Developmental needs: language expression skills development; Interest needs: strong interest in animal stories; Emotional needs: need for security and intimacy; Social needs: learning to express and listen; Cognitive needs: understanding the storyline and character relationships;

[0139] Parents’ educational style: Authoritativeness: 0.4 (medium); Responsiveness: 0.8 (high); Control: 0.3 (low); Support: 0.9 (high); Education preference: Focus on interaction and discussion;

[0140] Role assignment results: (Language education expert, reducer, 0.4): Provide reading skills related to language development; (Emotional connection consultant, remotional, 0.3): Enhance the emotional connection between parent-child reading; (Creative storyteller, rcreative, 0.2): Provide creative storytelling methods; (Reading interest cultivation teacher, rplay, 0.1): Design interesting reading extension activities;

[0141] 2. Outdoor exploration scene:

[0142] Interaction context: Interaction type: outdoor activities; Interaction location: park; Interaction activity: natural exploration; Interaction goal: scientific cognition, physical activity; Historical interaction: less outdoor exploration experience;

[0143] Children's needs: Developmental needs: development of gross motor skills and observation ability; Interest needs: interest in insects and plants; Emotional needs: seeking novel experiences and a sense of accomplishment; Social needs: relatively low, more focused on exploration; Cognitive needs: understanding natural phenomena and biological characteristics;

[0144] Parents’ educational style: Authoritativeness: 0.6 (medium-high); Responsiveness: 0.7 (medium-high); Control: 0.5 (medium); Support: 0.8 (high); Educational preference: Focus on practice and experience;

[0145] Role allocation results: (Natural Science Educator, reducer, 0.35): Provide scientific knowledge and exploration methods; (Safety Guardian, rsafety, 0.25): Ensure the safety of outdoor activities; (Exploration Guide, rplay, 0.25): Design interesting exploration activities; (Parent Coach, rparent, 0.15): Guide parents on how to promote learning in outdoor activities;

[0146] 3. Emotional management scenarios:

[0147] Interaction context: Interaction type: Emotional education; Interaction location: Family; Interaction activity: Emotional recognition and management training; Interaction goal: Cultivate emotion recognition and self-regulation abilities; Historical interaction: Children have had multiple experiences of emotional outbursts;

[0148] Children's needs: Developmental needs: Develop emotional regulation skills; Interest needs: Curiosity about expressions and body language; Emotional needs: A sense of security that requires understanding and acceptance; Social needs: Learning how to express emotions appropriately; Cognitive needs: Understanding the relationship between emotions and behavior;

[0149] Parental Education Style: Authoritativeness: 0.5 (medium); Responsiveness: 0.6 (medium); Control: 0.7 (medium-high); Support: 0.6 (medium); Education Preference: Focus on rules and behavior management;

[0150] Role allocation results: (Emotional education expert, remotional, 0.4): Provide professional guidance on emotion recognition and expression; (Behavior management consultant, reducer, 0.3): Design education programs for emotion and behavior management; (Parent coach, rparent, 0.2): Guide parents on how to deal with children's emotional problems; (Social skills trainer, rsocial, 0.1): Provide appropriate ways to express emotions in social situations.

[0151] The above embodiments show that the intelligent agent role allocation model can accurately match role combinations according to scene differences through multi-dimensional feature fusion and dynamic weight calculation, significantly improving the targetedness and collaborative efficiency of parent-child interaction.

[0152] In this embodiment, the effectiveness of the agent role allocation model needs to be ensured through continuous evaluation and optimization. Specifically,

[0153] The effectiveness of role allocation can be evaluated using the following indicators:

[0154] AllocationEffectiveness=α·InteractionQuality+β·DevelopmentProgress+γ1·UserSatisfaction;

[0155] Among them, InteractionQuality∈[0,1] represents the interaction quality score, which is based on the quality of the interaction process and results;

[0156] DevelopmentProgress∈[0,1] represents the developmental progress score, based on the child’s progress in the target ability;

[0157] UserSatisfaction∈[0,1] represents the user satisfaction score based on the feedback from parents and children;

[0158] α, β, and γ1 all represent weight coefficients, satisfying α+β+γ1=1.

[0159] Furthermore, it also includes:

[0160] The agent role allocation model is optimized according to the following formula:

[0161] ft+1=ft+η·∇fAllocationEffectiveness(ft);

[0162] Among them, ft+1 represents the role allocation function of the t+1th iteration; ft represents the role allocation function of the tth iteration; η represents the learning rate, which is used to control the optimization step size; ∇fAllocationEffectiveness(ft) represents the gradient of the effect score with respect to the role allocation function of the tth iteration.

[0163] Furthermore, optimization can be performed at different levels: for example, feature extraction can be optimized to improve the method of extracting features from interaction situations, children's needs and parents' educational styles; or, matching functions can be optimized to adjust the matching algorithm between roles and situations; or, weight distribution can be optimized to optimize the weight distribution strategy in multi-role combinations.

[0164] In some embodiments, this embodiment is designed from three aspects: response speed optimization, cold start strategy and privacy security, targeting the system integration requirements of the intelligent agent role allocation model, so as to ensure the efficiency and security of the system.

[0165] In terms of response speed optimization, the feature caching mechanism is used to reduce repeated calculation overhead. For example, feature vectors of high-frequency interaction scenarios (such as reading and outdoor activities) are cached to avoid re-extraction each time they are called. At the same time, a hierarchical calculation strategy is adopted. First, coarse-grained screening is performed (such as quickly eliminating unmatched roles based on interaction type), and then the candidate roles are refined for fitness calculation to reduce the computational complexity. In addition, through parallel processing technology (such as multi-threaded calculation of the matching degree of different roles), the time consumption of role allocation is significantly shortened. For example, in the reading scenario, the system can calculate the fitness of "language education experts" and "emotional connection consultants" in parallel to improve the overall response efficiency.

[0166] In terms of cold-start strategies, for new users or scenarios, we combine rule-based initial assignment (for example, pre-defining "Language Education Expert" as the default role for reading scenarios) with similar user migration (for example, referencing role assignment experience based on users with similar historical reading behavior). We also introduce an active exploration mechanism, trying multiple role combinations in the initial phase (for example, activating both "Creative Storyteller" and "Reading Interest Cultivator" simultaneously). This allows for rapid calibration of role weights using short-term feedback data, shortening the cold-start period.

[0167] In terms of privacy and security measures, data privacy is protected through data anonymization (such as deleting user IDs and encrypting sensitive fields); local computing strategies calculate key features (such as children's emotion recognition results) on the terminal device to avoid uploading raw data; or, access control mechanisms limit the permissions of different role agents (such as "security guardians" can only access location information) to prevent unauthorized access; and regular security audits (such as reviewing the data flow path in the role assignment logic) ensure that the system complies with privacy protection standards.

[0168] Through the above method, the intelligent agent role allocation model can dynamically allocate the most suitable combination of intelligent agent roles according to the interaction situation, children's needs and parents' educational styles, provide precise support for each parent-child interaction scenario, and significantly improve the personalization level and educational effect of the system.

[0169] S103: Input multiple intelligent agents and the weights of each intelligent agent into the intelligent agent collaborative decision-making model to obtain interaction decision-making information;

[0170] Specifically, inputting multiple intelligent agents and the weights of each intelligent agent into the intelligent agent collaborative decision-making model to obtain interaction decision-making information includes:

[0171] Obtain the constraint conditions for decision generation;

[0172] Generate multiple candidate decisions according to the constraint conditions, interaction situation data, children's needs data and parents' educational style data;

[0173] Calculate the weighted total utility scores of each candidate decision according to multiple intelligent agents and the weights of each intelligent agent;

[0174] Select the candidate decision with the highest weighted total utility score as the interaction decision-making information.

[0175] By combining the intelligent agent weights with the candidate decision attributes, the system realizes multi-dimensional utility evaluation. This mechanism ensures that the decision not only meets the core goals (such as language development) but also takes into account secondary needs (such as emotional connection). At the same time, by sorting the weighted total utility scores, the scheme with the highest comprehensive score is selected from multiple candidate decisions, avoiding the local optimum problem caused by single-dimensional optimization.

[0176] In this embodiment, the intelligent agent collaborative decision-making model is the core component in the intelligent agent driving engine, which is used to integrate the decision-making suggestions of multiple professional intelligent agents and generate parent-child interaction plans. This model balances various considerations by weighted aggregation of the utility evaluations of different intelligent agents, and provides comprehensive and precise guidance for parent-child interaction. This mathematical expression enables the system to simulate the collaborative decision-making process of an expert team and provide multi-angle professional support for family education.

[0177] Specifically, the mathematical expression of the intelligent agent collaborative decision-making model is as follows:

[0178] ;

[0179] Among them, D* ∈ D is the interaction decision-making information, representing the recommended parent-child interaction plan; D represents the space of all possible decisions.

[0180] The interaction decision-making information can be represented as a structured data:

[0181] ;

[0182] Among them, aD represents interactive activities, such as reading, games, conversations and other specific activity types; cD represents interactive content, such as specific books, game rules, conversation topics, etc.; rD represents interactive roles, such as parental roles, participation methods and responsibilities of virtual characters; gD represents interactive goals, such as cultivating specific abilities, solving specific problems, and enhancing emotional connections.

[0183] D∈Dempty is a candidate decision, representing a possible interaction plan considered by the system.

[0184] In practical applications, the system will generate multiple candidate decisions {D1, D2, ..., Dm};

[0185] Candidate decisions can be obtained through template generation, historical memory retrieval, or agent creative generation;

[0186] Each candidate decision needs to meet basic feasibility constraints, such as age suitability, safety, and resource availability.

[0187] Furthermore, n represents the number of agents involved in collaborative decision-making. This typically includes agents with diverse professional roles, such as educational experts, psychological counselors, and game designers. The number of agents can be dynamically adjusted based on the complexity and importance of the decision. Simple scenarios may only require 3-5 agents, while complex ones may require more than 10.

[0188] Furthermore, i∈{1, 2, ..., n} is an agent index, representing a specific professional agent. Each index corresponds to an agent with a specific professional background and responsibilities. Agents can have different professional fields, such as cognitive development, emotional education, and language learning. Agents can also have different roles, such as observer, analyst, creator, evaluator, and so on.

[0189] Furthermore, Ui(D)∈[0,1] is a utility function, which represents the evaluation score of the i-th agent on the decision D. The utility function reflects the agent's evaluation of the decision from its professional perspective.

[0190] In some embodiments, calculating the weighted total utility of each candidate decision based on multiple agents and the weight of each agent includes:

[0191] The utility of each agent for the candidate decision is calculated according to the following formula to obtain the utility score of each agent:

[0192] Ui(D)=fi(vi1(D),vi2(D),...vij(D)...,vik(D));

[0193] Among them, Ui(D) represents the utility score; vij(D) represents the score of the candidate decision on the j-th evaluation dimension that the i-th agent focuses on; fi represents the utility aggregation function of the i-th agent, such as weighted average, geometric mean, or more complex nonlinear functions.

[0194] Then, the utility score of each agent is multiplied by the weight of each agent to obtain the weighted total utility of the candidate decision.

[0195] Common evaluation dimensions include:

[0196] Developmental suitability: vdev(D)=sim(D, DevStage(c)), which is used to assess the degree of fit between the decision and the child's current developmental stage; DevStage(c) represents the developmental stage characteristics of child c; sim represents the similarity or fit function;

[0197] Educational effectiveness: vedu(D)=E[LearningGain(c, D)], used to evaluate the learning benefits that a decision may bring; LearningGain(c, D) represents the learning benefits that child c obtains from decision D; E[·] represents the expected value;

[0198] Emotional value: vemo(D) = EmotionalValue(D, EmoState(c, p)), which evaluates the emotional value and emotional connection promotion of the decision; EmoState(c, p) represents the current emotional state of child c and parent p; EmotionalValue represents the emotional value evaluation function;

[0199] Implementation feasibility: vfeas(D) = Feasibility(D, Resources), which is used to evaluate the implementation feasibility of the decision under current conditions; Resources represents available resources, such as time, materials, and space; Feasibility represents the feasibility evaluation function;

[0200] Long-term effects: , used to assess the long-term impact and sustainability of a decision. Impact(D, t) represents the impact of decision D after time t. γ∈(0, 1) is the discount factor. T is the time span considered.

[0201] Furthermore, wi∈[0, 1] is the weight coefficient, which indicates the importance of the i-th agent in decision making.

[0202] The weight coefficient satisfies the normalization condition: ;

[0203] Weights can be determined based on a variety of factors:

[0204] ;

[0205] Among them, Relevancei∈[0,1] represents the relevance between the agent’s professional field and the current decision-making problem; Confidencei∈[0,1] represents the agent’s confidence in its own evaluation; Performancei∈[0,1] represents the agent’s performance score in historical decision-making.

[0206] Weights can also be adjusted dynamically, based on the decision context and historical performance:

[0207] ;

[0208] in, represents the weight of agent i in the t-th decision; r(t) represents the actual effect score of the t-th decision; represents the average effect score of historical decisions; Ui(D(t)) represents the score of agent i on the tth decision; represents the average score of all agents for the tth decision; η represents the learning rate, which controls the speed of weight adjustment.

[0209] Furthermore, argmaxD is a maximization operation, which means selecting the decision that maximizes the sum of weighted utilities from all candidate decisions. In actual implementation, various optimization algorithms can be used to solve the maximization problem. For discrete decision spaces, the optimal solution can be found through enumeration or heuristic search. For continuous decision spaces, methods such as gradient ascent or evolutionary algorithms can be used.

[0210] To avoid decision-making homogeneity and ensure innovation, the system adopts the following methods to promote the diversity of interactive decision-making information:

[0211] ;

[0212] Among them, Diversity(D, HistoricalDecisions) represents the diversity measure of decision D and historical decisions, and λ represents the diversity weight, which is used to control the importance of diversity in decision making.

[0213] Diversity can be calculated as follows:

[0214] Where Distance(D, D′) represents the distance or difference measure between decisions D and D′.

[0215] In this embodiment, a specialized innovative agent is introduced, whose utility function focuses on evaluating the innovation and uniqueness of decisions:

[0216] Uinnovation(D)=Novelty(D)·Value(D);

[0217] Novelty(D) represents the novelty of decision D, which can be calculated based on its difference from known decisions. Value(D) represents the potential value of decision D, ensuring that the innovation is meaningful rather than random variation.

[0218] The following is an example of the application of the intelligent collaborative decision-making model in parent-child interaction scenarios:

[0219] 1. Personalized reading activity design:

[0220] Decision-making problem: Designing a reading activity for a 5-year-old child who is interested in dinosaurs but has a short attention span;

[0221] Participating agents: Reading education expert agent, used to evaluate the educational value and difficulty appropriateness of reading materials; child psychologist agent, used to evaluate the effectiveness of activities in cultivating attention; interest development consultant agent, used to evaluate the integration of activities with dinosaur interests; interaction designer agent, used to evaluate the fun and participation of activities; parent role consultant agent, used to evaluate the role positioning of parents in activities and the feasibility of implementation.

[0222] Collaborative decision-making process:

[0223] Generate candidate decisions: including interactive dinosaur encyclopedia reading, dinosaur story role-playing, and dinosaur illustration book production;

[0224] Agent evaluation: Each agent evaluates each candidate decision from its own professional perspective;

[0225] Weight allocation: Based on the characteristics of the problem, higher weights are given to attention- and interest-related agents;

[0226] Optimal decision: Choose the activity with the highest weighted utility and "Segmented Dinosaur Story Creation";

[0227] Decision-making outputs: including activity design, implementation steps, expected results and parent guidance suggestions;

[0228] 2. Parent-child conflict mediation plan:

[0229] Decision-making problem: Designing an intervention plan for a 3-year-old child's outburst in a public place;

[0230] Participating agents: Emotion Regulation Expert Agent, used to evaluate the effectiveness of the program on emotion regulation; Behavior Analyst Agent, used to analyze behavioral triggers and intervention points; Public Etiquette Educator Agent, used to evaluate the appropriateness of social norms education; Parent Stress Management Consultant Agent, used to evaluate the program's consideration of parental stress; Long-term Development Consultant Agent, used to evaluate the program's long-term educational value.

[0231] Collaborative decision-making process:

[0232] Generate candidate decisions: including preventive strategies, immediate intervention methods, follow-up education plans, etc.;

[0233] Agent evaluation: Each agent evaluates each candidate solution from different perspectives;

[0234] Weight allocation: Based on the urgency of the situation, higher weights are given to emotion regulation and behavior analysis agents;

[0235] Optimal decision: Choose a multi-stage approach that integrates emotion recognition, environmental adjustment, and positive guidance;

[0236] Decision-making outputs: including preventive measures, on-site intervention steps, follow-up education recommendations, and parent self-regulation guidance;

[0237] 3. Development Milestone Promotion Plan:

[0238] Decision-making problem: Designing a language development program for 2-year-old children approaching the critical period of language development;

[0239] Participating agents: Language development expert agent, used to evaluate the scientific nature of the language development plan; daily integration consultant agent, used to evaluate the feasibility of the plan in daily life; multi-sensory learning designer agent, used to evaluate the integration of multi-sensory stimulation; family resource assessor agent, used to evaluate the match between the plan and family resources; personalized education expert agent: evaluates the match between the plan and children's personality characteristics.

[0240] Collaborative decision-making process:

[0241] Generate candidate decisions: including structured language games, daily conversation reinforcement, multimedia assistance and other solutions;

[0242] Agent evaluation: Each agent evaluates each candidate plan from a professional perspective;

[0243] Weight distribution: Based on the characteristics of the critical development period, higher weights are given to language development and personalized education agents;

[0244] The best decision: Choose the "Contextual Language Interaction" program to integrate language learning into everyday situations;

[0245] Decision-making outputs: include daily activity suggestions, a list of key vocabulary, methods for monitoring progress, and adjustment mechanisms.

[0246] In this way, the intelligent collaborative decision-making model can integrate the knowledge and judgment of multiple professional intelligent agents to provide a comprehensive, balanced and personalized optimal solution for parent-child interaction, significantly improving the scientificity and effectiveness of family education.

[0247] S104: presenting the interactive decision information to the parent and the child, and obtaining interactive feedback information from the parent and the child after the interaction according to the interactive decision;

[0248] S105: Input the interactive feedback information into a loop iterative optimization algorithm to obtain updated interactive decision information.

[0249] Specifically, the interactive feedback information is input into the iterative optimization algorithm to obtain updated interactive decision information including:

[0250] Initialize the historical data set and set the initial prior distribution of the parameters;

[0251] At each step of the interaction, observe the current state;

[0252] Sample parameters from the current posterior distribution;

[0253] Calculate the action value of each action based on the sampled parameters and the action value function;

[0254] Select the action that maximizes the action value and obtain the target action;

[0255] Perform the target action and get the reward of feedback;

[0256] Combine the current state, target action, and reward to get a triple;

[0257] Add triples to the historical dataset for update;

[0258] Based on the updated historical data set, the posterior distribution is updated using the Bayesian formula;

[0259] Calculate the expected integral of each action based on the updated posterior distribution and the action value of each action;

[0260] The action with the highest expected score is selected as the updated interactive decision information.

[0261] More specifically, initialize the historical dataset D 数据 = {s1, a1, r1}, and set an initial prior distribution p0(θ) for the parameter θ. For example, in a parent-child reading scenario, the parameters might include the language education expert weight weducator and the emotional connection advisor weight wemotional. The initial prior distribution can be set based on domain knowledge (such as a Gaussian distribution with a mean of 0.5). The prior distribution p0(θ) is typically a conjugate prior to simplify calculations.

[0262] Then, at each interaction step, the current state st is observed (e.g., the child's emotion recognition result, the parent's voice command content), and the parameter θt is obtained by sampling from the posterior distribution. For example, if the current posterior distribution is N(μ=0.6, σ2=0.1), the sampling result is θt=0.63.

[0263] The system then calculates the expected value of each candidate action (such as "Increase the frequency of questioning" or "Introduce role-playing") using the action-value function Q(st, a) = ϕ(st, a)Tθt based on the sampled parameter θt. For example, if the feature vector ϕ of "Increase the frequency of questioning" is [1, 0.8], then its action value is Q = 1 × 0.63 + 0.8 × 0.63 = 1.134.

[0264] Then, the action at that maximizes the action value is selected as the target action, and the action is performed (such as actually increasing the frequency of asking questions) to obtain the feedback reward rt (such as the child's mood improvement score of 0.9).

[0265] Then the current state st, target action at and reward rt are combined into a triple (st, at, rt) and added to the historical dataset D 数据 , complete the data update.

[0266] Based on the updated historical dataset D 数据 , the posterior distribution p(θ|D) is updated using the Bayesian formula. For example, if new data shows that "increasing the frequency of asking questions" is effective in language development scenarios, the mean μ of the posterior distribution will shift to a higher value, reflecting an increase in confidence in the parameter θ.

[0267] Finally, based on the updated posterior distribution and action-value function, the expected score of all actions is calculated, and the action with the highest expected score is selected as the updated interactive decision information. For example, if the expected score of "Introducing Role Playing" is 1.2 (higher than other actions), then this action becomes the recommended strategy for the next stage.

[0268] This embodiment achieves continuous optimization of interactive decision-making through dynamic data collection and adaptive parameter updating, ensuring that the system can adjust strategies based on real-time feedback and adapt to complex and changing parent-child interaction scenarios.

[0269] In this embodiment, a Thompson sampling-based iterative optimization algorithm is the core algorithm in the agent-driven parent-child interaction loop system. It is used to balance exploration and exploitation in uncertain environments and continuously optimize the agent's decision-making strategy. This algorithm uses Bayesian inference to sample from the posterior distribution of the action-value function to achieve adaptive decision optimization. This mathematical formulation enables the system to gradually improve the quality of interaction suggestions with limited interaction data, providing personalized and continuously optimized parent-child interaction support.

[0270] Specifically, the mathematical expression of the loop iterative optimization algorithm is as follows:

[0271] ;

[0272] in, is the updated interactive decision information, indicating the optimal action that the agent should choose in state s; represents the space of all possible actions; in the parent-child interaction system, the action can be a recommended interactive activity, a dialogue strategy, or an educational intervention; the updated interactive decision information is a mapping that maps the state space to the action space, i.e. .

[0273] Furthermore, s∈S is the system state, representing the current context of parent-child interaction; S represents the space of all possible states.

[0274] The state can be represented as a structured data:

[0275] s=(cs,ps,es,hs);

[0276] Among them, cs represents child status, including age, development stage, interests, emotions, etc.; ps represents parent status, including educational style, available time, emotional state, etc.; es represents environment status, including scene, available resources, time, etc.; hs represents historical status, including summary information of past interactions.

[0277] Furthermore, a∈ Actions represent the decisions that the agent can take.

[0278] An action can be represented as a structured data:

[0279] a=(ta,ca,ma,ga);

[0280] Among them, ta represents the type of interaction, such as reading, games, conversation, creation, etc.; ca represents the content of interaction, such as specific books, game rules, conversation topics, etc.; ma represents the way of interaction, such as parental guidance, virtual character participation, etc.; ga represents the goal of interaction, such as cultivating specific abilities, solving specific problems, etc.

[0281] Further, is the action value function, which represents the expected cumulative reward of taking action a in state s. The action value function reflects the long-term value of the action, taking into account the immediate reward and the possible rewards in the future. In the parent-child interaction system, the reward can be the child's development progress, the improvement of the parent-child relationship, the improvement of the interaction quality, etc.

[0282] In some embodiments, calculating the action value of each action based on the sampled parameters using the action value function includes:

[0283] The action value of each action is calculated based on the sampling parameters and according to the following formula:

[0284] ;

[0285] Among them, Q(s,a) represents the action value of taking action a in state s; E represents the expectation function; rt+1 represents the reward obtained at time step t+1; γ∈[0,1] represents the discount factor; s0 represents the initial state; π represents the interactive decision information followed after the initial state action a0 pairs (s,a).

[0286] Furthermore, D is the historical data, which represents the set of state-action-reward triplets collected by the system; D={(s1, a1, r1), (s2, a2, r2), ..., (sn, an, rn)}; each triple (si, ai, ri) represents an observation of taking action ai in state si and obtaining reward ri; historical data is used to update the posterior distribution of the action-value function.

[0287] Furthermore, p(Q|D) is the posterior distribution, which represents the probability distribution of the action-value function Q given the historical data D. The posterior distribution is calculated from the prior distribution p(Q) and the likelihood function p(D|Q) via Bayesian updating:

[0288] p(Q|D)∝p(D|Q)·p(Q);

[0289] In practical applications, a specific parameterized form of the posterior distribution is usually assumed, such as a Gaussian distribution or a Beta distribution.

[0290] For linear function approximation, it can be assumed that:

[0291] ;

[0292] in, Feature vector representing a state-action pair; represents the parameter vector; the posterior distribution can be expressed as the distribution of parameter θ: p(θ|D).

[0293] Further, is the expected integral, which represents the expectation of the action value under the posterior distribution; this integral calculates the expected value of action a under the current information; in Thompson sampling, this expectation is not calculated directly, but a specific Q function is sampled from the posterior distribution; for the parameterized Q function, this is equivalent to sampling a parameter vector θ from the parameter posterior distribution p(θ|D).

[0294] Furthermore, argmaxa is a maximization operation, which means selecting the action that maximizes the expected action value. In Thompson sampling, this step is simplified to selecting the action that maximizes the sampled Q function value. This randomized maximization operation naturally balances exploration and utilization.

[0295] In this embodiment, for complex models, an approximate method can be used to update the posterior distribution:

[0296] Specifically, sampled variational inference can be performed:

[0297] ;

[0298] where Q is the family of approximate distributions and KL represents the Kullback-Leibler divergence.

[0299] Markov Chain Monte Carlo (MCMC) can also be used: by constructing a Markov chain, a sample sequence that obeys the posterior distribution is generated.

[0300] Particle filtering can also be used: a set of weighted particles is used to approximate the posterior distribution, and the particle weights are updated by resampling.

[0301] In some embodiments, Thompson sampling naturally balances exploration and exploitation. Its mechanism can be understood from the following perspectives:

[0302] 1. Uncertainty-driven exploration: For state-action pairs with less data, the variance of the posterior distribution is larger, resulting in increased variability in the sampled values. This increases the probability of selecting these uncertain actions and promotes exploration of unknown areas. As data accumulates, the posterior distribution gradually converges, and exploration naturally decreases.

[0303] 2. Probability matching: The probability of Thompson sampling selecting action a is equal to the probability that the action is optimal:

[0304] ;

[0305] This probability matching property makes the degree of exploration proportional to the uncertainty that an action might be optimal;

[0306] 3. Asymptotic optimality: As the number of interactions increases, the posterior distribution converges to the true action-value function, which ensures that the algorithm will eventually converge to the optimal strategy.

[0307] In the parent-child interactive virtual universe system, the specific applications of Thompson sampling strategy optimization include:

[0308] 1. Optimization of interactive activity recommendations:

[0309] Status representation: including children’s age, interests, ability level, parents’ educational style, etc.;

[0310] Action space: different types of parent-child interactive activities, such as reading, games, creation, etc.

[0311] Reward definition: based on indicators such as interaction quality, children's engagement, and learning outcomes;

[0312] Optimization process: The system initially sets priors based on expert knowledge, and gradually personalizes the recommendation strategy as family usage accumulates data;

[0313] 2. Dialogue strategy optimization:

[0314] State representation: including conversation history, child’s language ability, current topic, etc.;

[0315] Action space: different dialogue strategies, such as asking questions, explaining, guiding, expanding, etc.

[0316] Reward definition: based on indicators such as conversation continuity, quality of children’s expressions, and knowledge acquisition;

[0317] Optimization process: Systematically learning when to use which conversational strategies best promote children's language development;

[0318] 3. Optimization of virtual character behavior:

[0319] Status representation: including the current parent-child interaction status, child emotions, parental involvement, etc.;

[0320] Action space: different actions that the avatar can take, such as telling a story, providing knowledge, organizing a game, etc.

[0321] Definition of reward: Based on indicators such as the degree of harmony in the interaction between the three parties and the degree of achievement of educational goals;

[0322] Optimization process: The system learns when the virtual character should actively participate and when it should retreat to a supporting role;

[0323] 4. Optimization of educational intervention strategies:

[0324] Status representation: including children’s learning difficulties, learning style, emotional state, etc.;

[0325] Action space: different educational intervention strategies, such as direct instruction, guided discovery, and providing demonstrations;

[0326] Reward definition: Based on indicators such as learning outcomes, children’s self-confidence, and problem-solving abilities;

[0327] Optimization process: Systematically learn which intervention strategies are most effective for specific learning difficulties.

[0328] Through the above methods, agent strategy optimization based on Thompson sampling can balance exploration and utilization with limited interaction data, continuously improve the quality of parent-child interaction suggestions, and provide personalized and continuously optimized educational support for each family.

[0329] This embodiment achieves continuous optimization of parent-child interaction suggestions through the collaborative work of multiple agent roles. By introducing an agent role allocation model, the roles and functions of agents can be dynamically adjusted according to different interaction needs, thereby providing more precise support for parent-child interaction.

[0330] See also Figure 2 This embodiment provides a parent-child interaction decision information generating device 200, comprising:

[0331] The data acquisition unit 201 is used to acquire interaction context data, children's needs data and parents' education style data;

[0332] An agent acquisition unit 202 is configured to input the interaction context data, child demand data, and parental education style data into an agent role allocation model to obtain multiple agents and the weight of each agent;

[0333] A decision acquisition unit 203 is used to input multiple agents and the weight of each agent into the agent collaborative decision model to obtain interactive decision information;

[0334] A display unit 204 is used to display the interactive decision information to the parent and the child, and obtain interactive feedback information after the parent and the child interact according to the interactive decision;

[0335] The updating unit 205 is configured to input the interactive feedback information into a cyclic iterative optimization algorithm to obtain updated interactive decision information.

[0336] Furthermore, the agent acquisition unit 202 includes:

[0337] a feature extraction subunit, configured to extract features from the interaction context data, child demand data, and parent education style data, and generate interaction context features, child demand features, and parent education style features, respectively;

[0338] an integration subunit, configured to integrate the interaction context features, child demand features, and parental education style features into a comprehensive feature vector through a feature fusion function;

[0339] a scoring subunit, configured to calculate a fitness score for each role in a predefined role set based on the comprehensive feature vector;

[0340] A sorting subunit, for sorting the roles according to the compatibility score of each role and selecting compatible role combinations;

[0341] The normalization processing subunit is used to normalize the fitness scores of the selected roles and generate the weights of each role;

[0342] The allocation subunit is used to allocate a corresponding agent to each role according to the weight of each role and through the agent selection function to obtain multiple agents and the weight of each agent.

[0343] Furthermore, it also includes:

[0344] The optimization unit is used to optimize the agent role allocation model according to the following formula:

[0345] ;

[0346] Where ft+1 represents the role allocation function of the t+1th iteration; ft represents the role allocation function of the tth iteration; η represents the learning rate; represents the gradient of the effect score with respect to the role assignment function at the t-th iteration.

[0347] Furthermore, the decision acquisition unit 203 includes:

[0348] The condition acquisition subunit is used to obtain the constraint conditions for decision generation;

[0349] a candidate decision generating subunit, configured to generate a plurality of candidate decisions based on the constraint conditions, the interaction context data, the child's needs data, and the parenting style data;

[0350] The scoring calculation subunit is used to calculate the weighted total utility score of each candidate decision based on multiple agents and the weight of each agent;

[0351] The selection subunit is used to select the candidate decision with the highest weighted total utility score as the interactive decision information.

[0352] Furthermore, the score calculation subunit includes:

[0353] The utility score calculation subunit is used to calculate the utility of each agent for the candidate decision according to the following formula to obtain the utility score of each agent:

[0354] Ui(D)=fi(vi1(D),vi2(D),...vij(D)...,vik(D));

[0355] Where Ui(D) represents the utility score; vij(D) represents the score of the candidate decision on the jth evaluation dimension that the i-th agent focuses on; fi represents the utility aggregation function of the i-th agent;

[0356] The multiplication subunit is used to multiply the utility score of each agent and the weight of each agent to obtain the weighted total utility of the candidate decision.

[0357] Furthermore, the updating unit 205 includes:

[0358] Initialization subunit, used to initialize the historical data set and set the initial prior distribution of parameters;

[0359] The observation subunit is used to observe the current status at each step of the interaction;

[0360] The sampling subunit is used to sample parameters from the current posterior distribution;

[0361] A value calculation subunit, configured to calculate the action value of each action based on the sampled parameters and through the action value function;

[0362] The action selection subunit is used to select the action that maximizes the action value and obtain the target action;

[0363] An execution subunit, configured to execute the target action and obtain feedback;

[0364] a combining subunit, configured to combine the current state, the target action, and the reward to obtain a triple;

[0365] an adding subunit, configured to add the triple to the historical data set for updating;

[0366] The distribution updating subunit is used to update the posterior distribution through the Bayesian formula based on the updated historical data set;

[0367] The integral calculation subunit is used to calculate the expected integral of each action based on the updated posterior distribution and the action value of each action;

[0368] The information updating subunit is used to select the action with the highest expected score as the updated interactive decision information.

[0369] Furthermore, the value calculation subunit includes:

[0370] The action calculation subunit is used to calculate the action value of each action based on the sampling parameters and according to the following formula:

[0371] ;

[0372] Among them, Q(s, a) represents the action value of taking action a in state s; E represents the expected function; r t+1 represents the reward obtained at time step t; γ∈[0,1] represents the discount factor; s0 represents the initial state; π represents the interactive decision information followed after the initial state action a0 pair (s, a).

[0373] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-mentioned devices and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0374] The present invention also provides a computer-readable storage medium having a computer program stored thereon. When executed, the computer program can implement the methods provided in the above embodiments. The storage medium may include a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, among other media capable of storing program code.

[0375] The present invention further provides a computer device that may include a memory and a processor. The memory stores a computer program, and the processor, when invoking the computer program in the memory, can implement the method provided in the above embodiment. Of course, the computer device may also include various network interfaces, a power supply, and other components.

[0376] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the various embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description. It should be pointed out that for ordinary technicians in this technical field, without departing from the principle of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

[0377] It should also be noted that, in this specification, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprising" or any other variations thereof are intended to cover non-exclusive.

[0378] Inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a..." does not preclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.

Claims

1. A method for generating parent-child interaction decision information, characterized in that: include: Obtain interaction context data, children's needs data, and parents' educational style data; Inputting the interaction context data, child demand data, and parental education style data into an agent role allocation model to obtain multiple agents and the weight of each agent; Input multiple agents and the weights of each agent into the agent collaborative decision-making model to obtain interactive decision information; Presenting the interactive decision information to the parent and the child, and obtaining interactive feedback information from the parent and the child after the interaction according to the interactive decision; Inputting the interactive feedback information into a cyclic iterative optimization algorithm to obtain updated interactive decision information; The interactive context data, child demand data, and parental education style data are input into the agent role allocation model to obtain multiple agents and the weight of each agent, including: Extracting features from the interaction context data, child demand data, and parent education style data to generate interaction context features, child demand features, and parent education style features, respectively; The interaction context features, children's demand features and parents' educational style features are integrated into a comprehensive feature vector through a feature fusion function; Calculating a fitness score for each role in a predefined role set based on the comprehensive feature vector; Sort the characters according to their compatibility score and select compatible character combinations; Normalize the fitness scores of the selected roles and generate the weights of each role; According to the weight of each role and through the agent selection function, a corresponding agent is assigned to each role to obtain multiple agents and the weight of each agent; Also includes: The agent role allocation model is optimized according to the following formula: ； Among them, f t+1 represents the role allocation function of the t+1th iteration; f t represents the role allocation function of the tth iteration; η represents the learning rate; represents the gradient of the effect score with respect to the role assignment function at the t-th iteration.

2. The method for generating parent-child interaction decision information according to claim 1, characterized in that: The step of inputting multiple agents and the weight of each agent into the agent collaborative decision model to obtain interactive decision information includes: Obtain the constraints for decision generation; generating a plurality of candidate decisions based on the constraints, the interaction context data, the child's needs data, and the parenting style data; Calculate the weighted total utility score of each candidate decision based on multiple agents and the weight of each agent; The candidate decision with the highest weighted total utility score is selected as the interactive decision information.

3. The method for generating parent-child interaction decision information according to claim 2, characterized in that: The step of calculating the weighted total utility of each candidate decision based on multiple agents and the weight of each agent includes: The utility of each agent for the candidate decision is calculated according to the following formula to obtain the utility score of each agent: U i (D)=f i (v i1 (D),v i2 (D),...v ij (D)...,v ik (D)); Among them, U i (D) indicates utility score; v ij (D) represents the score of the candidate decision based on the j-th evaluation dimension that the i-th agent focuses on; f i represents the utility aggregation function of the i-th agent; The utility score of each agent is multiplied by the weight of each agent to obtain the weighted total utility of the candidate decision.

4. The method for generating parent-child interaction decision information according to claim 1, characterized in that: Inputting the interactive feedback information into the cyclic iterative optimization algorithm to obtain updated interactive decision information includes: Initialize the historical data set and set the initial prior distribution of the parameters; At each step of the interaction, observe the current state; Sample parameters from the current posterior distribution; Calculate the action value of each action based on the sampled parameters and the action value function; Select the action that maximizes the action value and obtain the target action; Execute the target action and obtain feedback in return; Combining the current state, the target action, and the reward to obtain a triple; Adding the triple to the historical data set for updating; Based on the updated historical data set, the posterior distribution is updated using the Bayesian formula; Calculate the expected integral of each action based on the updated posterior distribution and the action value of each action; The action with the highest expected score is selected as the updated interactive decision information.

5. The method for generating parent-child interaction decision information according to claim 4, characterized in that: The action value of each action is calculated based on the sampling parameters and the action value function, including: The action value of each action is calculated based on the sampling parameters and according to the following formula: ； Among them, Q(s, a) represents the action value of taking action a in state s; E represents the expected function; r t+1 represents the reward obtained at time step t+1; γ∈[0,1] represents the discount factor; s0 represents the initial state; π represents the interactive decision information followed after the initial state action a0 pair (s, a).

6. A parent-child interaction decision information generation device, characterized in that: include: A data acquisition unit, used to acquire interaction context data, children's needs data, and parents' education style data; An agent acquisition unit, configured to input the interaction context data, child demand data, and parental education style data into an agent role allocation model to obtain multiple agents and the weight of each agent; A decision acquisition unit is used to input multiple agents and the weights of each agent into the agent collaborative decision model to obtain interactive decision information; A display unit, configured to display the interactive decision information to the parent and the child, and obtain interactive feedback information from the parent and the child after the interaction according to the interactive decision; An updating unit, configured to input the interactive feedback information into a cyclic iterative optimization algorithm to obtain updated interactive decision information; The agent acquisition unit includes: a feature extraction subunit, configured to extract features from the interaction context data, child demand data, and parent education style data, and generate interaction context features, child demand features, and parent education style features, respectively; an integration subunit, configured to integrate the interaction context features, child demand features, and parental education style features into a comprehensive feature vector through a feature fusion function; a scoring subunit, configured to calculate a fitness score for each role in a predefined role set based on the comprehensive feature vector; A sorting subunit, for sorting the roles according to the compatibility score of each role and selecting compatible role combinations; The normalization processing subunit is used to normalize the fitness scores of the selected roles and generate the weights of each role; The allocation subunit is used to allocate a corresponding agent to each role according to the weight of each role and through the agent selection function, thereby obtaining multiple agents and the weight of each agent; Also includes: The optimization unit is used to optimize the agent role allocation model according to the following formula: ； Among them, f t+1 represents the role allocation function of the t+1th iteration; f t represents the role allocation function of the t-th iteration; η represents the learning rate; ∇fAllocationEffectiveness(f t ) represents the gradient of the effect score with respect to the role assignment function at the t-th iteration.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, it implements the parent-child interaction decision information generating method according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, causes the processor to execute the parent-child interaction decision information generating method according to any one of claims 1 to 5.

Images