Intelligent platform base construction method for AI employees and digital avatars
Patent Information
- Application Number
- CN202610841389.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-11
- Publication Date
- 2026-08-28
Smart Images
Figure CN122655841A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for constructing an intelligent platform foundation for AI employees and digital avatars. Background Technology
[0002] With the rapid development of artificial intelligence technology, AI employees and digital avatars are widely used in various industries. AI employees refer to intelligent agents built based on artificial intelligence technology that can perform specific tasks, while digital avatars refer to virtual mappings of physical people or systems in digital space. In existing technologies, the construction of AI employees and digital avatars is usually carried out independently, and there is a lack of effective collaboration mechanisms between the two. Existing collaboration solutions mainly adopt data-level fusion methods, which simply splice or weightedly fuse the output data of AI employees and digital avatars. This method cannot solve the fundamental differences between the two at the cognitive level, leading to frequent problems such as misunderstanding of goals, decision conflicts, and uncoordinated action sequences during the collaboration process.
[0003] Existing technologies lack the ability to assess the long-term impact of decisions and cannot predict the propagation effect of current decisions in the causal chain. This leads to collaborative decision-making often focusing only on immediate effects while ignoring long-term consequences. Therefore, how to build an intelligent platform foundation that enables deep cognitive collaboration between AI employees and digital avatars and has causal reasoning capabilities has become an urgent problem to be solved. Summary of the Invention
[0004] This invention provides a method for constructing an intelligent platform foundation for AI employees and digital avatars, in order to solve the problems mentioned in the background art.
[0005] To achieve the above objectives, the present invention provides a method for constructing an intelligent platform foundation for AI employees and digital avatars, comprising: S1 collects cognitive state data of AI employees and operational cognitive data of digital clones respectively. The two types of heterogeneous cognitive data are mapped to a unified semantic alignment space through a cognitive semantic projection network to obtain a semantically aligned cognitive representation. S2, based on the semantically aligned cognitive representation, combined with temporal constraints and domain prior knowledge, the Bayesian causal discovery framework is used to identify the causal relationship between AI employee behavior and digital clone status, generating a cross-modal causal association graph; S3. Based on the cross-modal causal correlation graph, an adversarial co-evolutionary learning framework is designed to achieve the co-evolution of dual-modal cognition under the premise of following causal constraints, and to obtain the co-evolutionary cognitive model. The adversarial co-evolutionary learning framework includes two adversarial networks, namely a consistency discriminator and a personalization generator. S4. Based on the co-evolved cognitive model and cross-modal causal relationship graph, a causal-driven decision fusion mechanism is constructed to predict the long-term impact of decisions through causal reasoning and generate intelligent collaborative decisions. S5. The semantically aligned cognitive representation, cross-modal causal association graph, co-evolved cognitive model, and causal-driven decision fusion mechanism are integrated into a unified platform to establish the foundation of the intelligent platform.
[0006] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention constructs a dual-mode cognitive representation and semantic alignment mechanism to map the heterogeneous cognitive data of AI employees and digital avatars to a unified semantic alignment space, solving the problem of the two being difficult to communicate at the cognitive level in the prior art. Based on the semantically aligned cognitive representation, AI employees and digital avatars can understand task goals and environmental states under a unified semantic framework, significantly reducing the goal comprehension bias and decision-making conflict in the collaboration process, and greatly improving collaboration efficiency.
[0007] 2. This invention achieves the co-evolution of AI employees and digital avatar cognitive models by constructing a cross-modal causal association graph and combining it with an adversarial co-evolutionary learning framework. It also has causal reasoning capabilities. Based on a causal-driven decision fusion mechanism, the system can predict the long-term impact of the current decision on the causal chain, so that the generated intelligent collaborative decision not only focuses on the immediate effect, but also takes into account the long-term consequences, and the quality of decision-making is significantly improved. The hierarchical architecture design of the unified platform base ensures the efficient integration and continuous optimization capabilities of each module. Attached Figure Description
[0008] Figure 1 This is a flowchart illustrating a method for constructing an intelligent platform foundation for AI employees and digital avatars, as provided in an embodiment of the present invention. The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0009] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0010] This application provides a method for constructing an intelligent platform foundation for AI employees and digital avatars. The execution entity of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the method for constructing an intelligent platform foundation for AI employees and digital avatars can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.
[0011] Reference Figure 1 The diagram shown is a flowchart illustrating a method for constructing an intelligent platform foundation for AI employees and digital avatars according to an embodiment of the present invention. In this embodiment, the method for constructing an intelligent platform foundation for AI employees and digital avatars includes: S1 collects cognitive state data of AI employees and operational cognitive data of digital clones respectively. The two types of heterogeneous cognitive data are mapped to a unified semantic alignment space through a cognitive semantic projection network to obtain a semantically aligned cognitive representation. In this embodiment of the invention, the step of collecting cognitive state data of the AI employee and operational cognitive data of the digital avatar respectively includes: Collect cognitive state data of AI employees, including task understanding representation, decision logic representation, behavior pattern representation and metacognitive monitoring data; Collect operational cognitive data of the digital clone, including state understanding representation, action logic representation, environmental perception representation, and performance self-evaluation data.
[0012] It should be noted that the task understanding representation in the cognitive state data of AI employees is obtained by extracting the semantic vector of the task description text through a natural language encoder; the decision logic representation is obtained by recording the key feature attention distribution in the decision-making process; the behavior pattern representation is obtained by analyzing the latent state features of the behavior sequence; and the metacognitive monitoring data is obtained by estimating the confidence level of its own decisions.
[0013] It should be noted that the state understanding representation in the operational cognitive data of the digital clone is a feature vector obtained by encoding the current environmental state through an environmental state encoder; the action logic representation is obtained by the action probability distribution output by the policy network; the environmental perception representation is a perception feature vector obtained by fusing data from multiple sensors; and the performance self-evaluation data is a quantitative indicator obtained by evaluating the task completion rate in real time.
[0014] It should be noted that the collection of these cognitive data adopted non-intrusive monitoring technology to ensure that the normal operation of AI employees and digital avatars is not affected; the extraction of task understanding representations converts natural language task descriptions into fixed-dimensional semantic vectors; the acquisition of decision logic representations reflects the key areas of focus in the decision-making process by recording the attention weight distribution of the intermediate layers of the neural network.
[0015] Furthermore, the collection of behavioral pattern representations uses sequence modeling techniques to encode historical behavioral sequences to obtain representation vectors; metacognitive monitoring data is generated through a confidence calibration network, which evaluates the reliability of current decisions; the state understanding representation of the digital avatar uses a visual or state encoder to extract features of the environmental state, and the action logic representation is obtained from the action distribution output by the policy network.
[0016] Furthermore, the generation of environmental perception representation involves multimodal sensor data fusion, including multiple perception modalities such as vision, hearing, and touch; performance self-evaluation data is calculated in real time through reward functions or target completion evaluation modules; all collected data has been standardized to ensure the uniformity of data scale.
[0017] Furthermore, the collection frequency of cognitive state data and operational cognitive data is dynamically adjusted according to task requirements; for tasks requiring rapid response, a high-frequency collection mode is adopted; for tasks requiring in-depth analysis, a lower-frequency but more detailed collection mode is adopted; a data quality detection mechanism is also implemented during the data collection process to filter out outliers and incomplete data.
[0018] In this embodiment of the invention, the step of mapping two types of heterogeneous cognitive data to a unified semantic alignment space through a cognitive semantic projection network to obtain a semantically aligned cognitive representation includes: A cognitive semantic projection network is designed using a hierarchical contrastive projection architecture, which includes a modality-specific encoder, a cross-modality attention layer, and a contrastive projection layer. The modality-specific encoder is used to encode the cognitive state data of the AI employee and the operational cognitive data of the digital clone to obtain the encoded feature representation. By establishing a bidirectional attentional association between the two types of cognitive data based on the encoded feature representation through the cross-modal attention layer, attention-enhanced feature representations are obtained; The attention-enhanced feature representation is projected onto a semantic alignment space of uniform dimension through the contrast projection layer to obtain a semantically aligned cognitive representation.
[0019] It should be noted that the hierarchical contrastive projection architecture processes AI employee cognitive data and digital clone cognitive data separately; each encoder contains a multi-head self-attention mechanism and a feedforward neural network, which can capture complex semantic relationships within their respective modalities.
[0020] Furthermore, the feature encoding process of the modality-specific encoder is as follows: First, the cognitive state data of the AI employee is input into the Transformer encoder to obtain the encoded feature representation; at the same time, the operational cognitive data of the digital clone is input into another Transformer encoder to obtain the corresponding encoded feature representation; the two encoders have the same structure but do not share parameters.
[0021] It should be noted that the bidirectional attention association establishment process of the cross-modal attention layer adopts a scaled dot product attention mechanism. This layer calculates the attention association from AI employee to digital avatar and from digital avatar to AI employee, establishing a bidirectional semantic connection. The calculation process of the cross-modal attention layer is as follows: First, the projections of the query vector, key vector, and value vector are calculated; then, the dot product of the query vector and key vector is calculated, divided by a scaling factor, and the softmax function is applied to obtain the attention weights; finally, the attention weights are multiplied by the value vector to obtain the attention output. This process calculates the attention associations in both directions separately. The mathematical expression of the cross-modal attention layer is as follows: Attention calculations from AI employees to digital avatars are as follows:
[0022] The attention calculation from digital avatars to AI employees is as follows:
[0023] In the formula, For attention, from AI employees to digital avatars For the attention of AI employees, from digital avatars to AI employees and These are the feature representations encoded from AI employee cognitive data and digital avatar cognitive data, respectively. , , For a learnable projection matrix, Let be the dimension of the key vector. This is the scaling factor.
[0024] It should be noted that the attention-enhanced feature representation is obtained by adding the original encoded features to the cross-modal attention output. The attention-enhanced features on the AI employee side are as follows: The attention enhancement feature of the digital clone side is .
[0025] Furthermore, the contrastive projection layer maps the high-dimensional attention-enhanced features to a unified-dimensional semantic alignment space, eliminating structural differences between modalities. The projection process of the contrastive projection layer is as follows: First, the encoded features are added to the cross-modal attention output to obtain the attention-enhanced features; then, the attention-enhanced features are subjected to layer normalization; finally, the normalized features are multiplied by the projection matrix to obtain the semantically aligned cognitive representation.
[0026] It should be noted that semantically aligned cognitions have the same dimension, which allows cognitions of different modalities to be compared and computed in the same semantic space.
[0027] Furthermore, the cognitive semantic projection network is trained using a contrastive learning objective. The contrastive loss function encourages the cognitive representations of AI employees and digital avatars under the same task to be close in the semantic alignment space, while the cognitive representations under different tasks are far apart in the semantic alignment space. The calculation process of the contrastive loss function is as follows: First, the similarity score of positive sample pairs (cognitive representations of AI employees and digital avatars under the same task) is calculated; then, the similarity score between the positive sample and all negative samples is calculated; next, the exponent of the positive sample similarity score is divided by the sum of the exponents of the similarity scores of all samples; finally, the negative logarithm is taken to obtain the contrastive loss value.
[0028] It should be noted that the temperature parameter controls the sensitivity of the similarity score; a smaller temperature parameter value makes the model pay more attention to difficult negative samples, while a larger temperature parameter value makes the similarity distribution smoother. The default value of the temperature parameter is set to 0.07, which can be modified by the operator according to the actual situation. The default value is the average of the most recent 10 temperature parameters each time the platform is rebuilt.
[0029] Furthermore, to prevent the loss of the characteristics of each modality during semantic alignment, the network also introduces an adversarial feature discriminator; the feature discriminator attempts to distinguish the source modality of the cognitive representation of semantic alignment, while the projection network attempts to generate a representation that deceives the discriminator; this adversarial training ensures that the unique characteristics of each modality are preserved while achieving semantic alignment.
[0030] It should be noted that the choice of semantic alignment space dimension is dynamically adjusted according to task complexity; 128 dimensions are used for simple tasks, and 512 dimensions are used for complex tasks.
[0031] S2, based on the semantically aligned cognitive representation, combined with temporal constraints and domain prior knowledge, the Bayesian causal discovery framework is used to identify the causal relationship between AI employee behavior and digital clone status, generating a cross-modal causal association graph; In this embodiment of the invention, the step of using a Bayesian causal discovery framework to identify the causal relationship between AI employee behavior and digital avatar status, and generating a cross-modal causal association graph, includes: Construct a Bayesian causal discovery framework that combines temporal constraints, domain prior knowledge, and sparse priors; Based on the semantically aligned cognitive representation, a Bayesian causal discovery framework is used to learn the causal structure and obtain the initial causal structure. The initial causal structure is refined and optimized to obtain the optimized causal structure; The optimized causal structure is subjected to causal effect quantification and uncertainty estimation to generate a cross-modal causal association map.
[0032] It should be noted that the Bayesian causal discovery framework adopts a hierarchical Bayesian modeling method; this framework transforms the causal discovery problem into a posterior probability inference problem, inferring the most likely causal structure by combining prior knowledge and observational data; the core of the framework is to calculate the posterior probability of different causal structures under given data using Bayes' theorem.
[0033] It should be noted that the optimization method transforms the discrete causal structure search problem into a continuous optimization problem. By introducing differentiable acyclic constraints, the causal structure can be optimized using gradient descent. The objective function of this optimization process is:
[0034] In the formula, This is a causal weight matrix. Denotes the Frobenius norm. Describing the L1 norm, is the sparsity regularization coefficient.
[0035] Furthermore, under the condition of satisfying the constraints Minimize the objective function under the condition of The function is an acyclic constraint function, and its expression is: ,in, This represents element-wise multiplication. Represents the trace of a matrix. For variable dimensions.
[0036] It should be noted that the acyclic constraint function ensures that the graph corresponding to the causal weight matrix is an acyclic directed graph; this constraint is implemented through a matrix exponential function, and the augmented Lagrangian method is used to handle the constraint if and only if the graph is acyclic.
[0037] Furthermore, causal effect quantification employs a do-calculus framework to calculate the average causal effect; for dependent variables... to variable The causal relationship, the average causal effect is defined as:
[0038] in, For variables to variable Average causal effect This indicates an intervention operation, which will involve variables. Forced to be set to value , This represents the expected value.
[0039] It should be noted that uncertainty estimation constructs multiple bootstrap sample sets by sampling with replacement from the original data multiple times. The causal discovery process is repeated on each sample set to obtain an estimate of the distribution of causal effects. The mathematical expression for uncertainty estimation is: Assuming to proceed After Bootstrap resampling, the following results were obtained One causal effect estimate Then the 95% confidence interval is: ,in, Indicates the first A series of ordinal statistics.
[0040] It should be noted that the generation of the cross-modal causal association graph includes the following steps: first, construct a directed acyclic graph based on the optimized causal structure; then, label the causal effect strength and confidence interval for each causal edge; finally, visualize the graph for easy understanding and analysis.
[0041] Furthermore, the cross-modal causal relationship graph contains three types of nodes: AI employee cognitive variable nodes, digital clone state variable nodes, and shared environmental variable nodes; edges represent causal relationships, the direction of the edges represents the causal direction, the thickness of the edges represents the causal strength, and the color of the edges represents the confidence level.
[0042] It should be noted that, for the problem of small sample size, the framework adopts transfer learning and data augmentation strategies; transfer learning uses causal knowledge of related tasks to initialize model parameters; data augmentation expands the training data by adding noise and generating synthetic samples.
[0043] Furthermore, for the problem of high-dimensional causal discovery, the framework adopts a hierarchical causal discovery strategy. The first layer uses a fast screening method to identify possible causal edges; the second layer uses a refined causal discovery method on the screened variables; and the third layer combines time series information to determine the causal direction.
[0044] It should be noted that the injection of domain prior knowledge is achieved through a constraint matrix; the known causal relationships provided by experts are encoded into a constraint matrix, where... Indicates from variable to variable The causal relationship is established. Indicates uncertainty. This indicates that the item definitely does not exist.
[0045] Furthermore, temporal constraints are implemented using time stamps. Each variable carries timestamp information, and the causal discovery process mandates that the timestamp of the causal variable be earlier than that of the outcome variable. For simultaneous variables, an instantaneous causal model is used.
[0046] It should be noted that the sparsity prior is achieved through L1 regularization. The L1 regularization coefficient controls the sparsity of the causal graph. Larger values produce a sparser causal graph. Cross-validation is used to select the model and balance the goodness of fit and simplicity.
[0047] Furthermore, applications of cross-modal causal relationship mapping include: identifying key impact paths of AI employee decisions on digital avatar states; discovering the reverse impact of digital avatar behavior on AI employee cognition; revealing key causal mechanisms in dual-modal collaboration; and providing structured constraints for collaborative evolution.
[0048] In this embodiment of the invention, the mathematical expression of the Bayesian causal discovery framework is as follows: ; In the formula, For the posterior probability distribution, For observed data, i.e., semantically aligned cognitive representations, It is a causal structure. Prior knowledge includes temporal constraints and domain knowledge. Given a causal structure Data likelihood To combine prior knowledge The prior probability distribution of the causal structure. Represents every possible causal structure. The total number of samples, and For sample index, The function is a Gaussian distribution. For variables The set of parent nodes, for, For the proportional sign, For noise variance, From variable to variable Causal weights.
[0049] It should be noted that prior probability is decomposed into three components: temporal constraint prior, domain knowledge prior, and sparsity prior. The temporal constraint prior forces the causal variable to precede the outcome variable in time; the domain knowledge prior injects causal relationships known to experts; and the sparsity prior encourages the generation of concise causal structures.
[0050] S3. Based on the cross-modal causal correlation graph, an adversarial co-evolutionary learning framework is designed to achieve the co-evolution of dual-modal cognition under the premise of following causal constraints, and to obtain the co-evolutionary cognitive model. The adversarial co-evolutionary learning framework includes two adversarial networks, namely a consistency discriminator and a personalization generator. In this embodiment of the invention, the step of designing an adversarial co-evolutionary learning framework based on the cross-modal causal correlation graph, and achieving co-evolution of bimodal cognition under the premise of following causal constraints, to obtain a co-evolved cognitive model, includes: An adversarial co-evolutionary learning framework is constructed, which includes a consistency discriminator and a personalization generator; The three levels of coordination and consistency are defined, including goal consistency, strategy complementarity, and temporal coordination. Based on the cross-modal causal relationship graph, causal constraint terms are designed. Through adversarial training of the consistency discriminator and the personalization generator, the co-evolution of dual-modal cognition is achieved under the premise of following the causal constraint terms, resulting in a co-evolved cognitive model. The causal constraint terms include causal path constraints and counterfactual consistency constraints. The co-evolved cognitive model includes an AI employee co-cognitive model and a digital clone co-cognitive model.
[0051] It should be noted that the adversarial co-evolutionary learning framework adopts a generative adversarial network architecture; the personalized generator is responsible for generating cognitive strategies that adapt to their respective environments, and the consistency discriminator is responsible for evaluating the co-consistency of the generated strategies; the two networks promote each other through adversarial training to achieve co-evolution of cognition.
[0052] Furthermore, the three levels of synergy consistency are specifically defined as follows: Goal consistency measures the degree to which the AI employee and the digital avatar have a consistent understanding of the task objectives; Strategy complementarity measures the degree to which the action strategies of the two complement each other rather than conflict; Timing coordination measures the degree of coordination and cooperation at the time of action; These three levels of synergy consistency ensure comprehensive synergy effects through multi-dimensional evaluation.
[0053] It should be noted that the design of causal constraints is based on cross-modal causal relationship graphs. Causal path constraints ensure that decisions do not violate known causal paths, and counterfactual consistency constraints ensure that decision logic remains consistent in counterfactual situations. These constraints integrate causal knowledge into the co-evolutionary process to ensure that the evolutionary direction conforms to the real causal relationship.
[0054] Furthermore, the adversarial training process employs an alternating optimization strategy. First, the generator parameters are fixed, and the discriminator parameters are updated to better distinguish between cooperative and inconsistent policy pairs. Then, the discriminator parameters are fixed, and the generator parameters are updated to generate policies that are both adapted to their respective environments and maintain cooperation. During the training process, the weight of causal constraints is gradually increased to ensure that the evolution process converges smoothly.
[0055] It should be noted that the co-evolved cognitive model consists of two parts: the AI employee collaborative cognitive model and the digital avatar collaborative cognitive model. These two models share some parameters to achieve knowledge sharing, while retaining independent parameters to adapt to their respective characteristics. The model adopts a deep neural network architecture, which can handle complex cognitive states and decision-making tasks.
[0056] Furthermore, the network architecture of the personalization generator adopts a dual-branch design. One branch processes the cognitive state and environmental information of the AI employee to generate the AI employee's collaborative strategy; the other branch processes the operational state and environmental information of the digital avatar to generate the digital avatar's collaborative strategy. The two branches share the feature extraction network of the intermediate layer to achieve knowledge transfer.
[0057] It should be noted that the network architecture of the consistency discriminator adopts a multi-head attention mechanism. The discriminator receives the policy pair, environment state and task description output by the generator as input, evaluates the collaborative consistency of different dimensions through the multi-head attention mechanism, and finally outputs a comprehensive collaborative consistency score.
[0058] Furthermore, the calculation of target consistency adopts the semantic similarity method. First, the encoded representations of the task target by the AI employee and the digital clone are extracted respectively. Then, the cosine similarity between the two encoded representations is calculated as the target consistency score. The encoded representations are obtained through a pre-trained language model or target encoding network.
[0059] It should be noted that the calculation of policy complementarity combines KL divergence and conflict detection. First, the KL divergence of the two policy distributions is calculated to measure the degree of difference. Then, a conflict detection function is used to identify direct conflicts between policies. Finally, the two are combined to obtain the policy complementarity score. The conflict detection function analyzes the degree of contradiction between policies at key decision points.
[0060] Furthermore, the calculation of temporal coordination employs a time difference penalty function. It records the time points of action of the AI employee and the digital avatar, calculates the absolute value of the time difference, and converts the time difference into a coordination score using an exponential decay function. The smaller the time difference, the higher the coordination score, encouraging time synchronization of actions.
[0061] Counterfactual strategies are obtained by intervening in certain variables in a causal graph, and measure the consistency of decision-making logic in different contexts.
[0062] It should be noted that the applications of the co-evolved cognitive model include: providing AI employees with collaborative decision-making suggestions that take into account the state of digital avatars; providing digital avatars with collaborative action guidance that takes into account the cognition of AI employees; supporting task planning and execution for dual-mode collaboration; and providing real-time evaluation and feedback on collaborative effects.
[0063] Furthermore, the training data for the model includes historical collaborative task records, single-model task data, and simulated collaborative scenario data. The data is cleaned and enhanced to ensure coverage of diverse collaborative scenarios and challenges. The training process adopts a course-based learning strategy, starting with simple collaborative tasks and gradually increasing the complexity of the tasks.
[0064] It should be noted that the monitoring and evaluation of the collaborative evolution process adopts a multi-indicator system, including indicators such as task completion rate, collaborative efficiency, conflict resolution rate, and user satisfaction. The training parameters and constraint weights are dynamically adjusted based on the evaluation results to achieve adaptive optimization.
[0065] Furthermore, the model deployment adopts an online learning and incremental update mechanism, continuously collecting collaborative data during practical applications, regularly updating model parameters, adapting to environmental changes and task evolution, supporting hot updates and version rollback, and ensuring the stability and reliability of the system.
[0066] In this embodiment of the invention, the mathematical expression of the adversarial co-evolutionary learning framework is as follows: ; In the formula, For the co-evolution objective function, This is the complete set of parameters for the AI employee cognitive model. This is the complete set of parameters for the digital clone cognitive model. To complete the loss function for collaborative tasks, This is the weighting coefficient for the consistency loss, with a value of 0.5. This is the weighting coefficient for the causal constraint loss, with a value of 0.3. For the collaborative consistency loss function, For causal constraint loss function, For the expectation operator, subscript Indicates from dataset Random sampling environment status and task description For each sample pair, take the average of all samples. For the training dataset, To provide decision-making strategies for AI employee cognitive models. The decision-making strategy output for the digital avatar cognitive model. A reference strategy vector for AI employees. The reference strategy vector for the digital clone. This is the weighting coefficient for goal consistency, with a value of 0.4. As a target consistency indicator, This is the weighting coefficient for strategic complementarity, with a value of 0.35. As a strategy complementarity indicator, This is a weighting coefficient for timing consistency, with a value of 0.25. As a time-series coordination indicator, This is the weighting coefficient for the causal path constraint, with a value of 0.05. For causal path constraint values, This is the weighting coefficient for the counterfactual consistency constraint, with a value of 0.03. For counterfactual consistency constraint values, The cosine similarity function is used. Let KL divergence function be used. For collision detection functions, This is a time scale parameter, with a value of 1 second. The threshold for permissible difference in influence is inversely proportional to the strength of the causal edge. To account for the differences in the impact of decisions on causal variables, This represents the actual decision-making strategy vector. This is the counterfactual decision strategy vector.
[0067] It should be noted that the training objective is to minimize the co-evolution objective function.
[0068] It should be noted that for tasks requiring a high level of understanding of the objectives, the weighting coefficient for objective consistency can be increased; for tasks requiring a high level of strategic coordination, the weighting coefficient for strategic complementarity can be increased; and for tasks requiring a high level of temporal synchronization, the weighting coefficient for temporal coordination can be increased.
[0069] It should be noted that operators can adjust the default values of all weight parameters according to the actual task requirements; each time the platform is rebuilt, the default value of the weight parameters is taken as the average value of the weight parameters used in the last 10 successful training sessions, so as to achieve adaptive optimization.
[0070] S4. Based on the co-evolved cognitive model and cross-modal causal relationship graph, a causal-driven decision fusion mechanism is constructed to predict the long-term impact of decisions through causal reasoning and generate intelligent collaborative decisions. In this embodiment of the invention, the construction of a causal-driven decision fusion mechanism, which predicts the long-term impact of decisions through causal reasoning and generates intelligent collaborative decisions, includes: Obtain the current environmental status, task description, and candidate decision-making strategies generated by the AI employee and digital avatar respectively; Based on the cross-modal causal association graph, causal intervention inference is performed on the candidate decision strategies, and the intervention effect value of each candidate decision strategy at each node of the causal association graph is calculated. Based on the intervention effect value, combined with the co-evolutionary cognitive model, the long-term impact assessment value of each candidate decision strategy is predicted based on the long-term impact assessment algorithm. Based on the aforementioned long-term impact assessment value, the candidate decision-making strategies for AI employees and digital avatars are integrated and optimized to generate intelligent collaborative decisions.
[0071] It should be noted that the causal-driven decision fusion mechanism is a bridge connecting the co-evolutionary cognitive model and the actual decision execution. The core of this mechanism is to use cross-modal causal association graphs to perform causal reasoning on candidate decisions, predict the propagation effect and long-term impact of decisions on the causal chain, and thus select the optimal co-decision scheme.
[0072] It should be noted that the candidate decision-making strategies generated by the AI employee and the digital avatar are generated by the co-evolved cognitive model based on the current environmental state and task description. Each candidate strategy corresponds to an action probability distribution, which represents the preference for selection on the set of available actions.
[0073] Furthermore, the causal intervention inference process is as follows: For each candidate decision strategy, firstly, the specific actions or decision variables contained in the strategy are taken as the intervention objects; then, based on the cross-modal causal association graph, the do-operator is used to calculate the causal effect of the intervention on each node in the graph, obtaining the intervention effect value vector. The formula for calculating the intervention effect value is... ,in This represents the decision variable that is being intervened upon. This indicates the value of the variable in the candidate strategy. Represents the first in the causal relationship graph Each node variable This represents the expected value.
[0074] It should be noted that the specific implementation steps of causal intervention inference are as follows: First, determine the set of intrusive variables involved in the candidate decision strategy; then, for each intrusive variable, set the corresponding node in the causal association graph as the intervention state; next, through the structural equation model of the causal association graph, propagate the intervention effect along the causal edge and calculate the variable change of all downstream nodes; finally, summarize to obtain the intervention effect value vector of the candidate strategy on all nodes.
[0075] It should be noted that the fusion optimization process is as follows: For the multiple candidate decision strategies generated by the AI employee and the digital avatar, the long-term impact assessment value of each candidate strategy is calculated; the candidate strategy with the highest assessment value is selected as the initial recommended strategy for each; then, it is checked whether there is a conflict between the two initial recommended strategies. If there is a conflict, it is fine-tuned through a negotiation adjustment mechanism until a conflict-free and optimal intelligent collaborative decision is generated.
[0076] Furthermore, the implementation steps of the negotiation adjustment mechanism are as follows: When there is a conflict between the initial recommendation strategies of AI employees and digital avatars, the specific type of conflict is first identified, such as resource conflict, action conflict, and timing conflict; then, based on the type of conflict, the causal path related to the conflict node is found in the causal relationship graph; next, the root cause of the conflict is traced back along the causal path; finally, based on the causal reasoning results, the conflicting actions in one or two strategies are adjusted so that the adjusted strategy can eliminate the conflict while maintaining a high evaluation value as much as possible.
[0077] It should be noted that the generated intelligent collaborative decision consists of three parts: the final decision-making strategy of the AI employee, the final decision-making strategy of the digital avatar, and the collaborative execution plan between the two. The collaborative execution plan includes the time sequence of action execution, resource allocation scheme, and contingency plan for handling abnormal situations, to ensure that the AI employee and the digital avatar can execute in an orderly manner according to the decision-making plan.
[0078] Furthermore, the causal-driven decision fusion mechanism supports online updates and adaptive optimization. After each intelligent collaborative decision is executed, the system collects actual feedback data during the execution process, including task completion status, actual observations of causal effects, and actual performance of collaborative consistency. This feedback data is used to update the causal strength parameters in the causal relationship graph and the parameters of the cognitive model, making subsequent decision fusion more accurate.
[0079] In this embodiment of the invention, the algorithm for the long-term impact assessment is as follows: ; In the formula, For long-term impact assessment values, This represents the immediate performance score of the candidate strategy. The causal reasoning score represents the candidate strategy. This represents the collaborative consistency score of the candidate strategies. , and These represent the weighting coefficients for the three scoring items, with default values of 0.3, 0.4, and 0.3 respectively.
[0080] S5. The semantically aligned cognitive representation, cross-modal causal association graph, co-evolved cognitive model, and causal-driven decision fusion mechanism are integrated into a unified platform to establish the foundation of the intelligent platform.
[0081] In this embodiment of the invention, the integration of the semantically aligned cognitive representation, the cross-modal causal association graph, the co-evolved cognitive model, and the causal-driven decision fusion mechanism into a unified platform to establish an intelligent platform foundation includes: A unified platform architecture is constructed, in which the semantically aligned cognitive representation is deployed on the cognitive alignment layer, the cross-modal causal association graph is deployed on the causal inference layer, and the co-evolved cognitive model and the causal-driven decision fusion mechanism are deployed on the collaborative decision layer. The unified platform architecture includes a data access layer, a cognitive alignment layer, a causal inference layer, a collaborative decision layer, and an application interface layer. The data access layer receives real-time environmental data and task data, which are then processed layer by layer by the cognitive alignment layer, causal reasoning layer, and collaborative decision-making layer. Finally, the application interface layer outputs intelligent collaborative decisions, thus establishing the foundation of the intelligent platform.
[0082] It should be noted that the unified platform architecture adopts a layered design. The data access layer is responsible for receiving and preprocessing real-time environmental data, task data, and cognitive state data from AI employees and digital avatars. The cognitive alignment layer carries the cognitive representation of the semantic alignment and performs semantic alignment processing on the input heterogeneous cognitive data. The causal reasoning layer carries the cross-modal causal association graph and performs causal analysis and reasoning on the aligned cognitive data. The collaborative decision-making layer carries the cognitive model after collaborative evolution and the causal-driven decision fusion mechanism to generate the final intelligent collaborative decision. The application interface layer outputs the generated intelligent collaborative decision to the AI employees and digital avatars for execution.
[0083] Furthermore, the layers interact with each other through standardized interfaces. The output of the data access layer serves as the input of the cognitive alignment layer, the output of the cognitive alignment layer serves as the input of the causal reasoning layer, the output of the causal reasoning layer serves as the input of the collaborative decision-making layer, and the output of the collaborative decision-making layer provides decision-making services to the outside world through the application interface layer. The layers support the forward propagation of data flow and the reverse propagation of feedback signals, realizing the platform's continuous optimization and self-evolution capabilities.
[0084] In the several embodiments provided by this invention, it should be understood that the disclosed method can be implemented in other ways.
[0085] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.
[0086] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, and technology that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.
[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.
Claims
1. A method for constructing an intelligent platform foundation for AI employees and digital avatars, characterized in that: The method includes: S1 collects cognitive state data of AI employees and operational cognitive data of digital clones respectively. The two types of heterogeneous cognitive data are mapped to a unified semantic alignment space through a cognitive semantic projection network to obtain a semantically aligned cognitive representation. S2, based on the semantically aligned cognitive representation, combined with temporal constraints and domain prior knowledge, the Bayesian causal discovery framework is used to identify the causal relationship between AI employee behavior and digital clone status, generating a cross-modal causal association graph; S3. Based on the cross-modal causal correlation graph, an adversarial co-evolutionary learning framework is designed to achieve the co-evolution of dual-modal cognition under the premise of following causal constraints, and to obtain the co-evolutionary cognitive model. The adversarial co-evolutionary learning framework includes two adversarial networks, namely a consistency discriminator and a personalization generator. S4. Based on the co-evolved cognitive model and cross-modal causal relationship graph, a causal-driven decision fusion mechanism is constructed to predict the long-term impact of decisions through causal reasoning and generate intelligent collaborative decisions. S5. The semantically aligned cognitive representation, cross-modal causal association graph, co-evolved cognitive model, and causal-driven decision fusion mechanism are integrated into a unified platform to establish the foundation of the intelligent platform.
2. The method for constructing an intelligent platform foundation for AI employees and digital avatars as described in claim 1, characterized in that, The collection of cognitive state data from AI employees and operational cognitive data from digital avatars includes: Collect cognitive state data of AI employees, including task understanding representation, decision logic representation, behavior pattern representation and metacognitive monitoring data; Collect operational cognitive data of the digital clone, including state understanding representation, action logic representation, environmental perception representation, and performance self-evaluation data.
3. The method for constructing an intelligent platform foundation for AI employees and digital avatars as described in claim 1, characterized in that, The process of mapping two types of heterogeneous cognitive data to a unified semantic alignment space through a cognitive semantic projection network to obtain a semantically aligned cognitive representation includes: A cognitive semantic projection network is designed using a hierarchical contrastive projection architecture, which includes a modality-specific encoder, a cross-modality attention layer, and a contrastive projection layer. The modality-specific encoder is used to encode the cognitive state data of the AI employee and the operational cognitive data of the digital clone to obtain the encoded feature representation. By establishing a bidirectional attentional association between the two types of cognitive data based on the encoded feature representation through the cross-modal attention layer, attention-enhanced feature representations are obtained; The attention-enhanced feature representation is projected onto a semantic alignment space of uniform dimension through the contrast projection layer to obtain a semantically aligned cognitive representation.
4. The method for constructing an intelligent platform foundation for AI employees and digital avatars as described in claim 1, characterized in that, The method of using a Bayesian causal discovery framework to identify the causal relationship between AI employee behavior and digital avatar status, and generating a cross-modal causal association graph, includes: Construct a Bayesian causal discovery framework that combines temporal constraints, domain prior knowledge, and sparse priors; Based on the semantically aligned cognitive representation, a Bayesian causal discovery framework is used to learn the causal structure and obtain the initial causal structure. The initial causal structure is refined and optimized to obtain the optimized causal structure; The optimized causal structure is subjected to causal effect quantification and uncertainty estimation to generate a cross-modal causal association map.
5. The method for constructing an intelligent platform foundation for AI employees and digital avatars as described in claim 4, characterized in that, The mathematical expression of the Bayesian causal discovery framework is as follows: ； In the formula, For the posterior probability distribution, For observed data, i.e., semantically aligned cognitive representations, It is a causal structure. Prior knowledge includes temporal constraints and domain knowledge. Given a causal structure Data likelihood To incorporate prior knowledge The prior probability distribution of the causal structure. Represents every possible causal structure. The total number of samples, and For sample index, The function is a Gaussian distribution. For variables The set of parent nodes, for, For the proportional sign, For noise variance, From variable to variable Causal weights.
6. The method for constructing an intelligent platform foundation for AI employees and digital avatars as described in claim 1, characterized in that, Based on the cross-modal causal correlation graph, an adversarial co-evolutionary learning framework is designed to achieve the co-evolution of bimodal cognition under the premise of following causal constraints, resulting in a co-evolutionary cognitive model, including: An adversarial co-evolutionary learning framework is constructed, which includes a consistency discriminator and a personalization generator; The three levels of coordination and consistency are defined, including goal consistency, strategy complementarity, and temporal coordination. Based on the cross-modal causal relationship graph, causal constraint terms are designed. Through adversarial training of the consistency discriminator and the personalization generator, the co-evolution of dual-modal cognition is achieved under the premise of following the causal constraint terms, resulting in a co-evolved cognitive model. The causal constraint terms include causal path constraints and counterfactual consistency constraints. The co-evolved cognitive model includes an AI employee co-cognitive model and a digital clone co-cognitive model.
7. The method for constructing an intelligent platform foundation for AI employees and digital avatars as described in claim 6, characterized in that, The mathematical expression for the adversarial co-evolutionary learning framework is as follows: ； In the formula, For the co-evolution objective function, This is the complete set of parameters for the AI employee cognitive model. This is the complete set of parameters for the digital clone cognitive model. To complete the loss function for collaborative tasks, This is the weighting coefficient for the consistency loss, with a value of 0.
5. This is the weighting coefficient for the causal constraint loss, with a value of 0.
3. For the collaborative consistency loss function, For causal constraint loss function, For the expectation operator, subscript Indicates from dataset Random sampling environment status and task description For each sample pair, take the average of all samples. For the training dataset, To provide decision-making strategies for AI employee cognitive models. The decision-making strategy output for the digital avatar cognitive model. A reference strategy vector for AI employees. The reference strategy vector for digital clones. This is the weighting coefficient for goal consistency, with a value of 0.
4. As a target consistency indicator, This is the weighting coefficient for strategic complementarity, with a value of 0.
35. As a strategy complementarity indicator, This is a weighting coefficient for timing consistency, with a value of 0.
25. As a time-series coordination indicator, This is the weighting coefficient for the causal path constraint, with a value of 0.
05. For causal path constraint values, This is the weighting coefficient for the counterfactual consistency constraint, with a value of 0.
03. For counterfactual consistency constraint values, The cosine similarity function is used. Let KL divergence function be used. For collision detection functions, This is a time scale parameter, with a value of 1 second. The threshold for permissible difference in influence is inversely proportional to the strength of the causal edge. To account for the differences in the impact of decisions on causal variables, This represents the actual decision-making strategy vector. This is the counterfactual decision strategy vector.
8. The method for constructing an intelligent platform foundation for AI employees and digital avatars as described in claim 1, characterized in that, The aforementioned causal-driven decision fusion mechanism, which predicts the long-term impact of decisions through causal reasoning and generates intelligent collaborative decisions, includes: Obtain the current environmental status, task description, and candidate decision-making strategies generated by the AI employee and digital avatar respectively; Based on the cross-modal causal association graph, causal intervention inference is performed on the candidate decision strategies, and the intervention effect value of each candidate decision strategy at each node of the causal association graph is calculated. Based on the intervention effect value, combined with the co-evolutionary cognitive model, the long-term impact assessment value of each candidate decision strategy is predicted based on the long-term impact assessment algorithm. Based on the aforementioned long-term impact assessment value, the candidate decision-making strategies for AI employees and digital avatars are integrated and optimized to generate intelligent collaborative decisions.
9. The method for constructing an intelligent platform foundation for AI employees and digital avatars as described in claim 8, characterized in that, The algorithm for the long-term impact assessment is as follows: ； In the formula, For long-term impact assessment values, This represents the immediate performance score of the candidate strategy. The causal reasoning score represents the candidate strategy. This represents the collaborative consistency score of the candidate strategies. , and These represent the weighting coefficients for the three scoring items, with default values of 0.3, 0.4, and 0.3 respectively.
10. The method for constructing an intelligent platform foundation for AI employees and digital avatars as described in claim 1, characterized in that, The integration of the semantically aligned cognitive representation, cross-modal causal association graph, co-evolved cognitive model, and causal-driven decision fusion mechanism into a unified platform to establish an intelligent platform foundation includes: A unified platform architecture is constructed, in which the semantically aligned cognitive representation is deployed on the cognitive alignment layer, the cross-modal causal association graph is deployed on the causal inference layer, and the co-evolved cognitive model and the causal-driven decision fusion mechanism are deployed on the collaborative decision layer. The unified platform architecture includes a data access layer, a cognitive alignment layer, a causal inference layer, a collaborative decision layer, and an application interface layer. The data access layer receives real-time environmental data and task data, which are then processed layer by layer by the cognitive alignment layer, causal reasoning layer, and collaborative decision-making layer. Finally, the application interface layer outputs intelligent collaborative decisions, thus establishing the foundation of the intelligent platform.