General artificial intelligence value function learning prediction method and system

Abstract

The invention provides a universal artificial intelligence value function learning prediction method and system. The method comprises the following steps: collecting character characteristic data of users through a multi-modal multi-user parallel cloud service platform, and obtaining observation data; analyzing graph coding through a graph machine learning model, and extracting effective structure graph information through a bottom-up and top-down combined attention mechanism model; intelligently learning a stable characteristic flow state from the observation data according to the effective structure diagram information, and obtaining the value of a sub-analysis diagram; establishing a comparative learning objective function model through the comparison of partial sequence data according to the value of the sub-analytic graph; minimizing the contrast learning objective function model, and constructing a multi-level value prediction function model; and multi-level value dimension prediction is carried out, and value-driven autonomous intelligent prediction is realized.

Description

Technical Field

[0001] The present invention relates to the field of reinforcement intelligence learning value prediction technology, and more specifically, to a general artificial intelligence value function learning prediction method and system. Background Art

[0002] At present, a large number of value function calculation works focus on how to learn the value function in reinforcement learning, learning the value function for a specific task, rather than learning the human value function. For example, generalized value function (GVFs) is a concept in reinforcement learning, which is used to express more types of knowledge by extending the traditional value function. In classical reinforcement learning, the value function is usually used to estimate the expected value of the future reward of the agent in a given state, while GVFs introduces more flexibility, allowing the definition of a variety of different goals or problem types to capture a wider range of knowledge structures; the current value alignment is to achieve consistency between human values ​​and the output value of the large model through examples; value alignment is a core concept in the field of artificial intelligence, which refers to making the behavior and decision of the agent consistent with human values, ethics and goals. Simply put, the purpose of value alignment is to prevent artificial intelligence systems from deviating from human expectations when learning and performing tasks autonomously, thereby causing potential dangers or negative effects; for strong artificial intelligence and complex AI systems that are widely used in human life, it is crucial to align them with human values ​​so that they can understand and perform tasks that are consistent with human values; but problems such as the lack of explicit construction of value functions remain to be solved; therefore, it is necessary to propose a general artificial intelligence value function learning prediction method and system to at least partially solve the problems existing in the existing technology. Summary of the invention

[0003] A series of simplified concepts are introduced in the content of the invention, which will be further described in detail in the specific implementation method section; the content of the invention of the present invention does not mean to attempt to limit the key features and essential technical features of the technical solution claimed for protection, nor does it mean to attempt to determine the scope of protection of the technical solution claimed for protection.

[0004] To at least partially solve the above problems, the present invention provides a general artificial intelligence value function learning prediction method, comprising: S100, collects user personality characteristic data and obtains observation data through a multi-modal multi-user parallel cloud service platform; S200, uses a graph machine learning model to parse graph encoding and extracts effective structural graph information through a combined attention mechanism model combining bottom-up and top-down methods; S300, based on the effective structure graph information, intelligently learns the stable dynamic characteristic flow state from the observation data to obtain the value of the sub-analysis graph; S400, according to the value of the sub-analysis graph, establishes a comparative learning objective function model through the comparison of partial order data; minimizes the comparative learning objective function model and constructs a multi-level value prediction function model; performs multi-level value dimension prediction to realize value-driven autonomous intelligent prediction.

[0005] Preferably, S100 includes: S101, setting a multi-modal multi-user parallel cloud service platform; a user accesses the multi-modal multi-user parallel cloud service platform through a network task interface; S102, displaying task scenarios in parallel through multiple network task interfaces; and providing task scenario collection prompts, prompting to set task operations, and obtaining task scenario sequences; S103, extracting the parsing graph sequence corresponding to the task scenario sequence.

[0006] Preferably, S200 includes: S201, based on the text representation of the parsing graph nodes and the edge features between nodes, the parsing graph is encoded through the graph machine learning model to obtain the node text feature vector and the edge text feature vector between nodes; S202, constructing a combined attention mechanism model by combining the bottom-up attention mechanism with the top-down attention mechanism; selecting the key areas of attention through the combined attention mechanism model, and extracting the key sub-parse graph and effective structure graph information.

[0007] Based on the text representation of the parsing graph nodes and the edge features between nodes, the parsing graph is encoded through the graph machine learning model to obtain the node text feature vector and the node edge text feature vector, including: parsing the graph based on the text representation of the parsing graph nodes and the edge features between nodes; obtaining the parsing graph node text features and the node edge text features according to the parsing graph node and the node edge feature vector; the text features cannot be directly involved in the calculation; the parsing graph is encoded through the graph machine learning model, the parsing graph node text features and the node edge text features are encoded respectively, the node text features and the node edge text features description format are unified, the text vector is involved in the calculation, and the node text feature vector and the node edge text feature vector are obtained; By combining the bottom-up attention mechanism with the top-down attention mechanism, a combined attention mechanism model is constructed; through the combined attention mechanism model, the key focus area is selected, and the key focus sub-parse graph and effective structure graph information are extracted, including: By combining the bottom-up attention mechanism with the top-down attention mechanism, a combined attention mechanism model is constructed, including: comparing the difference between the front and back frames of the parse graph sequence to obtain the front and back frame difference sequence; selecting the largest difference part in the front and back frame difference sequence as the focus part of the parse graph; combining the bottom-up attention mechanism with the top-down attention mechanism, according to the focus part of the parse graph, using a large language model to determine the key part with the highest correlation with the value dimension in the task scenario, thereby determining the key area from top to bottom; fusing the focus part with the key area to obtain the focus key area; constructing a combined attention mechanism model; through the combined attention mechanism model, extracting sub-parse graphs according to the focus key area to obtain the focus key sub-parse graph; selecting effective structure graph information according to the focus key sub-parse graph.

[0008] Preferably, S300 includes: S301, extracting local nodes in the effective structure graph information through a hybrid multi-head attention model according to the effective structure graph information; S302, according to the local node, obtain the sub-dimension value through the interactive neural network model; obtain the sub-dimension weight through the gate neural network; S303, intelligently learn the stable dynamic characteristic flow state from the observed data, perform weighted summation of the sub-dimension value and the sub-dimension weight, obtain the value of the value function, and obtain the value of the sub-analysis graph.

[0009] Preferably, S400 includes: S401, establishing a comparative learning objective function model by comparing the partial order data according to the value of the sub-analysis graph; S402, minimizing the contrast learning objective function model and calculating the value model parameters; S403, construct a multi-level value prediction function model based on the value model parameters; perform multi-level value dimension prediction based on the multi-level value prediction function model to achieve value-driven autonomous intelligent prediction.

[0010] According to the value model parameters, a multi-level value prediction function model is constructed; according to the multi-level value prediction function model, a multi-level value dimension prediction is performed to realize value-driven autonomous intelligent prediction, including: extracting the analysis graphs of two moments from the analysis graph sequence to obtain the analysis graph of the first moment and the analysis graph of the second moment; calculating the value of the first sub-analysis graph of the analysis graph at the first moment and the value of the second sub-analysis graph; calculating the value difference between the value of the first sub-analysis graph and the value of the second sub-analysis graph, and adding an adaptive factor of the interval between two moments to establish a comparative learning objective function model; minimizing the comparative learning objective function model and adding a mean sparse penalty term; the mean sparse penalty term is the product of the sparse factor and the weight mean; forming a multi-level value prediction function model; according to the multi-level value prediction function model, a multi-level value dimension prediction of the user group is performed to realize value-driven autonomous intelligent prediction.

[0011] The present invention provides a general artificial intelligence value function learning prediction system, comprising: The observation data feature cloud collection subsystem collects user personality feature data and obtains observation data through a multi-modal multi-user parallel cloud service platform; The graph model encoding extraction subsystem uses a graph machine learning model to parse graph encoding and extracts effective structural graph information through a combined attention mechanism model that combines bottom-up and top-down methods. The stable characteristic flow state learning subsystem intelligently learns the stable dynamic characteristic flow state from the observed data based on the effective structure diagram information and obtains the value of the sub-analysis diagram; The partial order comparative value prediction subsystem establishes a comparative learning objective function model based on the value of the sub-analysis graph and by comparing the partial order data; minimizes the comparative learning objective function model and constructs a multi-level value prediction function model; performs multi-level value dimension prediction and realizes value-driven autonomous intelligent prediction.

[0012] Preferably, the observation data feature cloud acquisition subsystem includes: A multi-modal multi-user parallel cloud interface module is used to set up a multi-modal multi-user parallel cloud service platform; users access the multi-modal multi-user parallel cloud service platform through a network task interface; The task interface parallel display module displays task scenarios in parallel through multiple network task interfaces; and provides task scenario collection prompts, prompts to set task operations, and obtains task scenario sequences; The scene parsing graph sequence extraction module extracts the parsing graph sequence corresponding to the task scene sequence.

[0013] Preferably, the stable characteristic flow state learning subsystem includes: The parsing graph encoding module encodes the parsing graph nodes and inter-node edge features based on the text representation, and obtains the node text feature vector and the inter-node edge text feature vector through the graph machine learning model; The combined attention mechanism extraction module constructs a combined attention mechanism model by combining the bottom-up attention mechanism with the top-down attention mechanism. Through the combined attention mechanism model, the key areas of attention are selected, and the key sub-parse graphs and effective structure graph information are extracted. Based on the text representation of the parsing graph nodes and the edge features between nodes, the parsing graph is encoded through the graph machine learning model to obtain the node text feature vector and the node edge text feature vector, including: parsing the graph based on the text representation of the parsing graph nodes and the edge features between nodes; obtaining the parsing graph node text features and the node edge text features according to the parsing graph node and the node edge feature vector; the text features cannot be directly involved in the calculation; the parsing graph is encoded through the graph machine learning model, the parsing graph node text features and the node edge text features are encoded respectively, the node text features and the node edge text features description format are unified, the text vector is involved in the calculation, and the node text feature vector and the node edge text feature vector are obtained; By combining the bottom-up attention mechanism with the top-down attention mechanism, a combined attention mechanism model is constructed; through the combined attention mechanism model, the key focus area is selected, and the key focus sub-parse graph and effective structure graph information are extracted, including: By combining the bottom-up attention mechanism with the top-down attention mechanism, a combined attention mechanism model is constructed, including: comparing the difference between the front and back frames of the parse graph sequence to obtain the front and back frame difference sequence; selecting the largest difference part in the front and back frame difference sequence as the focus part of the parse graph; combining the bottom-up attention mechanism with the top-down attention mechanism, according to the focus part of the parse graph, using a large language model to determine the key part with the highest correlation with the value dimension in the task scenario, thereby determining the key area from top to bottom; fusing the focus part with the key area to obtain the focus key area; constructing a combined attention mechanism model; through the combined attention mechanism model, extracting sub-parse graphs according to the focus key area to obtain the focus key sub-parse graph; selecting effective structure graph information according to the focus key sub-parse graph.

[0014] Preferably, the stable characteristic flow state learning subsystem includes: The hybrid multi-head node extraction module extracts local nodes from the effective structure graph information through the hybrid multi-head attention model according to the effective structure graph information; The sub-dimension weight calculation module obtains the sub-dimension value through the interactive neural network model according to the local node; and obtains the sub-dimension weight through the gate neural network; The analysis graph value calculation module intelligently learns stable dynamic characteristic flow states from the observed data, performs weighted summation of sub-dimension values ​​and sub-dimension weights, obtains the value function value, and obtains the value of the sub-analysis graph.

[0015] Preferably, the partial order comparison value prediction subsystem includes: The contrastive learning objective function model module establishes the contrastive learning objective function model by comparing the partial order data according to the value of the sub-analysis graph; The value model parameter calculation module minimizes the contrast learning objective function model and calculates the value model parameters; The multi-level value prediction module constructs a multi-level value prediction function model based on the value model parameters; based on the multi-level value prediction function model, it performs multi-level value dimension prediction to achieve value-driven autonomous intelligent prediction.

[0016] According to the value model parameters, a multi-level value prediction function model is constructed; according to the multi-level value prediction function model, a multi-level value dimension prediction is performed to realize value-driven autonomous intelligent prediction, including: extracting the analysis graphs of two moments from the analysis graph sequence to obtain the analysis graph of the first moment and the analysis graph of the second moment; calculating the value of the first sub-analysis graph of the analysis graph at the first moment and the value of the second sub-analysis graph; calculating the value difference between the value of the first sub-analysis graph and the value of the second sub-analysis graph, and adding an adaptive factor of the interval between two moments to establish a comparative learning objective function model; minimizing the comparative learning objective function model and adding a mean sparse penalty term; the mean sparse penalty term is the product of the sparse factor and the weight mean; forming a multi-level value prediction function model; according to the multi-level value prediction function model, a multi-level value dimension prediction of the user group is performed to realize value-driven autonomous intelligent prediction.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects: The present invention provides a general artificial intelligence value function learning prediction method and system, which collects user personality feature data and obtains observation data through a multi-modal multi-user parallel cloud service platform; performs parsing graph encoding through a graph machine learning model, and extracts effective structure graph information through a combined attention mechanism model combining bottom-up and top-down; based on the effective structure graph information, intelligently learns stable dynamic feature flow states from the observation data to obtain the value of the sub-parsing graph; based on the value of the sub-parsing graph, establishes a comparative learning objective function model through the comparison of partial order data; minimizes the comparative learning objective function model and constructs a multi-level value prediction function model; performs multi-level value dimension prediction to realize value-driven autonomous intelligent prediction; and realizes value alignment including: preference learning: the AI ​​system gradually learns and internalizes the value behind these behaviors by observing human choices and preferences. Intrinsic safety mechanisms: Introduce intrinsic safety mechanisms or restrictions in the design of AI to prevent decisions from going against human interests; including imposing constraints and restrictions on behaviors that may have negative consequences; Learning feedback: Guide AI to learn aligned values ​​through human supervision and feedback; for example, use human feedback reinforcement learning to allow AI to rely on value judgments provided by humans when making decisions; Two-way alignment: That is, in the process of AI design and deployment, not only does AI adapt to human values, but humans can also understand and adapt to AI's reasoning mode, forming a two-way communication and understanding of values; the present invention can learn multi-level value dimensions of user groups, such as cleanliness, comfort, safety, pro-sociality and other value dimensions, establish a mapping of any event and state to value dimensions, establish a foundation for an autonomous intelligent system driven by value, and realize the prediction and estimation of multi-level dimensional values.

[0018] The general artificial intelligence value function learning prediction method and system described in the present invention, other advantages, objectives and features of the present invention will be reflected in part through the following description, and in part will be understood by technical personnel in the field through research and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 This is a diagram of an embodiment of a general artificial intelligence value function learning prediction system described in the present invention.

[0020] Figure 2 This is another embodiment diagram of a general artificial intelligence value function learning prediction system described in the present invention.

[0021] Figure 3 This is a diagram of an embodiment of a general artificial intelligence value function learning prediction system and method described in the present invention. DETAILED DESCRIPTION

[0022] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments so that those skilled in the art can implement it according to the instructions; as shown in the figure, the present invention provides a general artificial intelligence value function learning prediction method, including: S100, collects user personality characteristic data and obtains observation data through a multi-modal multi-user parallel cloud service platform; S200, uses a graph machine learning model to parse graph encoding and extracts effective structural graph information through a combined attention mechanism model combining bottom-up and top-down methods; S300, based on the effective structure graph information, intelligently learns the stable dynamic characteristic flow state from the observation data to obtain the value of the sub-analysis graph; S400, according to the value of the sub-analysis graph, establishes a comparative learning objective function model through the comparison of partial order data; minimizes the comparative learning objective function model and constructs a multi-level value prediction function model; performs multi-level value dimension prediction to realize value-driven autonomous intelligent prediction.

[0023] The principle and effect of the above technical solution are as follows: the present invention provides a general artificial intelligence value function learning and prediction method, including: collecting user personality feature data and obtaining observation data through a multimodal multi-user parallel cloud service platform; encoding the parsing graph through a graph machine learning model, and extracting effective structure graph information through a combined attention mechanism model combining bottom-up and top-down; based on the effective structure graph information, intelligently learning stable dynamic feature flow states from the observation data to obtain the value of the sub-parsing graph; based on the value of the sub-parsing graph, establishing a comparative learning objective function model through the comparison of partial order data; minimizing the comparative learning objective function model and constructing a multi-level value prediction function model; performing multi-level value dimension prediction to realize value-driven autonomous intelligent prediction; sub-PG represents the effective structure graph information identifier; LGP represents the large parsing graph, Large Parser Graph; fluents means flow; h means personality trait; MLP means mapping function; V means value function value; achieving value alignment includes: preference learning: AI system gradually learns and internalizes the values ​​behind these behaviors by observing human choices and preferences; intrinsic safety mechanism: introduce intrinsic safety mechanisms or restrictions in the design of AI to prevent decisions that violate human interests; including imposing constraints and restrictions on behaviors that may have negative consequences; learning feedback: guiding AI to learn aligned values ​​through human supervision and feedback; for example, using human feedback reinforcement learning allows AI to rely on value judgments provided by humans when making decisions; two-way alignment: that is, in the process of AI design and deployment, not only AI is adapted to human values, but also humans can understand and adapt to AI's reasoning methods, forming a two-way value communication and understanding; multi-modal multi-user parallel cloud service platform includes: according to the non-complete accuracy of human personality characteristics, through probabilistic statistics, using multi-faceted domain Region to control multi-user parallel processing, Share cloud service resources, allocate cloud service resources to multiple network task interfaces in parallel, perform multimodal fusion of multiple user personality characteristics, multiply the user personality characteristics by their weights and then superimpose them to calculate the multimodal fusion personality characteristics; collect user personality characteristic data based on the multimodal fusion personality characteristics; user personality characteristic weights are distributed according to the probability of universal personality characteristics in human population statistics, and the universal personality characteristic probability corresponds to the probability of personality characteristic distribution in the human population. The higher the probability of personality characteristic distribution in the human population, the higher the probability of universal personality characteristics; for example, in the community work group, the prosocial value dimension is greater than the non-prosocial value dimension, and the personality characteristics corresponding to the prosocial value dimension are greater than the personality characteristics corresponding to the non-prosocial value dimension; in the social security work group, the personality characteristics corresponding to the security value dimension are greater than the personality characteristics corresponding to the non-security value dimension, and also greater than the personality characteristics corresponding to the prosocial value dimension, so as to reduce security risks; solve the technical problem of infinite collection range; significantly improve the accuracy of data collection for the incomplete accuracy of human personality characteristics;The present invention can learn the multi-level value dimensions of user groups, such as cleanliness, comfort, safety, pro-sociality, etc., establish a mapping of any event and state to value dimensions, establish the foundation of an autonomous intelligent system based on value drive, and realize the prediction and estimation of multi-level dimensional values. ;

[0024] In one embodiment, S100 includes: S101, setting a multi-modal multi-user parallel cloud service platform; a user accesses the multi-modal multi-user parallel cloud service platform through a network task interface; S102, displaying task scenarios in parallel through multiple network task interfaces; and providing task scenario collection prompts, prompting to set task operations, and obtaining task scenario sequences; S103, extracting the parsing graph sequence corresponding to the task scenario sequence.

[0025] The principle and effect of the above technical solution are: setting up a multimodal multi-user parallel cloud service platform; users access the multimodal multi-user parallel cloud service platform through a network task interface; displaying task scenarios in parallel through multiple network task interfaces; and providing task scenario collection prompts, prompting to set task operations, and obtaining task scenario sequences; extracting the task scenario sequence corresponding to the parsing graph sequence.

[0026] In one embodiment, S200 includes: S201, based on the text representation of the parsing graph nodes and the edge features between nodes, the parsing graph is encoded through the graph machine learning model to obtain the node text feature vector and the edge text feature vector between nodes; S202, constructing a combined attention mechanism model by combining the bottom-up attention mechanism with the top-down attention mechanism; selecting the key areas of attention through the combined attention mechanism model, and extracting the key sub-parse graph and effective structure graph information.

[0027] The principle and effect of the above technical solution are: based on the text representation of the parsing graph nodes and the edge features between nodes, the parsing graph is encoded through the graph machine learning model to obtain the node text feature vector and the node edge text feature vector; by combining the bottom-up attention mechanism with the top-down attention mechanism, a combined attention mechanism model is constructed; through the combined attention mechanism model, the key areas of attention are selected, and the key sub-parsing graphs and effective structure graph information are extracted; based on the text representation of the parsing graph nodes and the edge features between nodes, the parsing graph is encoded through the graph machine learning model to obtain the node text feature vector and the node edge text feature vector, including: the parsing graph represents the parsing graph nodes and the edge features between nodes based on the text; the parsing graph node text features and the node edge text features are obtained according to the parsing graph nodes and the node edge feature vectors; the text features cannot be directly involved in the calculation; the parsing graph is encoded through the graph machine learning model, the parsing graph node text features and the node edge text features are encoded respectively, the node text features and the node edge text features description formats are unified, the text vector participates in the calculation, and the node text feature vector and the node edge text feature vector are obtained; By combining the bottom-up attention mechanism with the top-down attention mechanism, a combined attention mechanism model is constructed; through the combined attention mechanism model, the key focus area is selected, and the key focus sub-parse graph and effective structure graph information are extracted, including: By combining the bottom-up attention mechanism with the top-down attention mechanism, a combined attention mechanism model is constructed, including: comparing the difference between the front and back frames of the parse graph sequence to obtain the front and back frame difference sequence; selecting the largest difference part in the front and back frame difference sequence as the focus part of the parse graph; combining the bottom-up attention mechanism with the top-down attention mechanism, according to the focus part of the parse graph, using a large language model to determine the key part with the highest correlation with the value dimension in the task scenario, thereby determining the key area from top to bottom; fusing the focus part with the key area to obtain the focus key area; constructing a combined attention mechanism model; through the combined attention mechanism model, extracting sub-parse graphs according to the focus key area to obtain the focus key sub-parse graph; selecting effective structure graph information according to the focus key sub-parse graph.

[0028] In one embodiment, S300 includes: S301, extracting local nodes in the effective structure graph information through a hybrid multi-head attention model according to the effective structure graph information; S302, according to the local node, obtain the sub-dimension value through the interactive neural network model; obtain the sub-dimension weight through the gate neural network; S303, intelligently learn the stable dynamic characteristic flow state from the observed data, perform weighted summation of the sub-dimension value and the sub-dimension weight, obtain the value of the value function, and obtain the value of the sub-analysis graph.

[0029] The principle and effect of the above technical solution are: according to the effective structure graph information, local nodes in the effective structure graph information are extracted through a hybrid multi-head attention model; according to the local nodes, sub-dimension values ​​are obtained through an interactive neural network model; sub-dimension weights are obtained through a gate neural network; stable dynamic feature flow states are intelligently learned from the observed data, and the sub-dimension values ​​and sub-dimension weights are weightedly summed to obtain the value function value, thereby obtaining the value of the sub-analysis graph.

[0030] In one embodiment, S400 includes: S401, establishing a comparative learning objective function model by comparing the partial order data according to the value of the sub-analysis graph; S402, minimizing the contrast learning objective function model and calculating the value model parameters; S403, constructing a multi-level value prediction function model according to the value model parameters; performing multi-level value dimension prediction according to the multi-level value prediction function model to achieve value-driven autonomous intelligent prediction; According to the value model parameters, a multi-level value prediction function model is constructed; according to the multi-level value prediction function model, a multi-level value dimension prediction is performed to realize value-driven autonomous intelligent prediction, including: extracting the analysis graphs of two moments from the analysis graph sequence to obtain the analysis graph of the first moment and the analysis graph of the second moment; calculating the value of the first sub-analysis graph of the analysis graph at the first moment and the value of the second sub-analysis graph; calculating the value difference between the value of the first sub-analysis graph and the value of the second sub-analysis graph, and adding an adaptive factor of the interval between two moments to establish a comparative learning objective function model; minimizing the comparative learning objective function model and adding a mean sparse penalty term; the mean sparse penalty term is the product of the sparse factor and the weight mean; forming a multi-level value prediction function model; according to the multi-level value prediction function model, a multi-level value dimension prediction of the user group is performed to realize value-driven autonomous intelligent prediction.

[0031] The principle and effect of the above technical solution are: according to the value of the sub-analysis graph, by comparing the partial order data, a comparative learning objective function model is established; the comparative learning objective function model is minimized to calculate the value model parameters; according to the value model parameters, a multi-level value prediction function model is constructed; according to the multi-level value prediction function model, multi-level value dimension prediction is carried out to realize value-driven autonomous intelligent prediction; according to the value model parameters, a multi-level value prediction function model is constructed; according to the multi-level value prediction function model, multi-level value dimension prediction is carried out to realize value-driven autonomous intelligent prediction, including: extracting the analysis graphs of two moments from the analysis graph sequence to obtain the analysis graph of the first moment and the analysis graph of the second moment; calculating the value of the first sub-analysis graph of the analysis graph at the first moment and the value of the second sub-analysis graph; calculating the value difference between the value of the first sub-analysis graph and the value of the second sub-analysis graph, and adding the adaptive factor of the interval between the two moments to establish the comparative learning objective function model; minimizing the comparative learning objective function model and adding the mean sparse penalty term; the mean sparse penalty term is the product of the sparse factor and the weight mean; forming a multi-level value prediction function model; according to the multi-level value prediction function model, multi-level value dimension prediction of the user group is carried out to realize value-driven autonomous intelligent prediction.

[0032] The present invention provides a general artificial intelligence value function learning prediction system, comprising: The observation data feature cloud collection subsystem 100 collects user personality feature data and obtains observation data through the multi-modal multi-user parallel cloud service platform 101; The graph model encoding extraction subsystem uses a graph machine learning model to parse graph encoding and extracts effective structural graph information through a combined attention mechanism model that combines bottom-up and top-down methods. The stable characteristic flow state learning subsystem intelligently learns the stable dynamic characteristic flow state from the observed data based on the effective structure diagram information and obtains the value of the sub-analysis diagram; The partial order comparative value prediction subsystem establishes a comparative learning objective function model based on the value of the sub-analysis graph and by comparing the partial order data; minimizes the comparative learning objective function model and constructs a multi-level value prediction function model; performs multi-level value dimension prediction and realizes value-driven autonomous intelligent prediction.

[0033] The principle and effect of the above technical solution are as follows: the present invention provides a general artificial intelligence value function learning prediction system, including: an observation data feature cloud acquisition subsystem 100, which collects user personality feature data and obtains observation data through a multimodal multi-user parallel cloud service platform 101; a graph model encoding extraction subsystem 200, which performs parsing graph 211 encoding 211a through a graph machine learning model 201, and extracts effective structure graph information 221 through a combined attention mechanism model 202 combining bottom-up and top-down methods; a stable feature flow state learning subsystem 300, which intelligently learns stable dynamic feature flow states from observation data based on effective structure graph information to obtain the value of sub-parse graphs; a partial order comparison value prediction subsystem 400, which establishes a comparison learning objective function 401 based on the value of sub-parse graphs through the comparison of partial order data; minimizes the comparison learning objective function model and constructs a multi-level value prediction function model 402; performs multi-level value dimension prediction to realize value-driven autonomous intelligent prediction; sub-PG represents effective structure graph information identification; LGP represents large parse graph, Large Parser Graph; fluents means flow; h means personality trait; MLP means mapping function; V means value function value; achieving value alignment includes: preference learning: AI system gradually learns and internalizes the values ​​behind these behaviors by observing human choices and preferences; intrinsic safety mechanism: introduce intrinsic safety mechanisms or restrictions in the design of AI to prevent decisions that violate human interests; including imposing constraints and restrictions on behaviors that may have negative consequences; learning feedback: guide AI to learn aligned values ​​through human supervision and feedback; for example, using human feedback reinforcement learning allows AI to rely on value judgments provided by humans when making decisions; two-way alignment: that is, in the process of AI design and deployment, not only does AI adapt to human values, but humans can also understand and adapt to AI's reasoning method, forming a two-way value communication and understanding; multimodal multi-user parallel cloud The service platform includes: according to the non-complete accuracy of human personality characteristics, through probabilistic statistics, using multi-domain Region to control multi-user parallel processing, sharing cloud service resources, cloud service resources are distributed in parallel to multiple network task interfaces, multi-user personality characteristics are multimodally fused, user personality characteristic weights are multiplied by user personality characteristics and then superimposed on each other to calculate multimodal fusion personality characteristics; user personality characteristic data is collected according to multimodal fusion personality characteristics; user personality characteristic weights are distributed according to the probability of universal personality characteristics of human population statistics, and the universal personality characteristic probability corresponds to the probability of personality characteristic distribution in the human population. The higher the probability of personality characteristic distribution in the human population, the higher the probability of universal personality characteristics; for example, in community work groups, the prosocial value dimension is greater than the non-prosocial value dimension, and the personality characteristics corresponding to the prosocial value dimension are greater than the personality characteristics corresponding to the non-prosocial value dimension;In the social security work group, the personality characteristics corresponding to the safety value dimension are greater than the personality characteristics corresponding to the non-safety value dimension, and also greater than the personality characteristics corresponding to the pro-social value dimension, so as to reduce safety risks; the technical problem of infinite collection range is solved; the accuracy of the data collected for the incomplete accuracy of human personality characteristics is significantly improved; the present invention can learn the multi-level value dimensions of the user group, such as cleanliness, comfort, safety, pro-sociality and other value dimensions, establish the mapping of any event and state to the value dimension, establish the foundation of the autonomous intelligent system based on value drive, and realize the prediction and estimation of multi-level dimensional values. ;

[0034] In one embodiment, the observation data feature cloud acquisition subsystem includes: A multi-modal multi-user parallel cloud interface module is used to set up a multi-modal multi-user parallel cloud service platform; users access the multi-modal multi-user parallel cloud service platform through a network task interface; The task interface parallel display module displays task scenarios in parallel through multiple network task interfaces; and provides task scenario collection prompts, prompts to set task operations, and obtains task scenario sequences; The scene parsing graph sequence extraction module extracts the parsing graph sequence corresponding to the task scene sequence.

[0035] The principle and effect of the above technical solution are: an observation data feature cloud acquisition subsystem, including: a multimodal multi-user parallel cloud interface module, setting a multimodal multi-user parallel cloud service platform; users access the multimodal multi-user parallel cloud service platform through a network task interface; a task interface parallel display module, which displays task scenarios in parallel through multiple network task interfaces; and performs task scenario acquisition prompts, prompting to set task operations and obtain task scenario sequences; a scene parsing graph sequence extraction module, which extracts the parsing graph sequence corresponding to the task scenario sequence.

[0036] In one embodiment, the graph model encoding extraction subsystem includes: The parsing graph encoding module encodes the parsing graph nodes and inter-node edge features based on the text representation, and obtains the node text feature vector and the inter-node edge text feature vector through the graph machine learning model; The combined attention mechanism extraction module constructs a combined attention mechanism model by combining the bottom-up attention mechanism with the top-down attention mechanism. Through the combined attention mechanism model, the key areas of attention are selected, and the key sub-parse graphs and effective structure graph information are extracted.

[0037] The principle and effect of the above technical solution are as follows: a graph model encoding extraction subsystem, including: a parsing graph encoding module, which performs parsing graph encoding based on text representation of parsing graph nodes and inter-node edge features through a graph machine learning model 201 to obtain node text feature vectors and inter-node edge text feature vectors; a combined attention mechanism extraction module, which combines the bottom-up attention mechanism with the top-down attention mechanism to construct a combined attention mechanism model 202; through the combined attention mechanism model, select key areas of attention, and extract key sub-parsing graphs 2020 and effective structure graph information 2021; based on the text representation of parsing graph nodes and inter-node edge features , the parsing graph is encoded by the graph machine learning model, and the node text feature vector and the node edge text feature vector are obtained, including: the parsing graph is represented based on the text to parse the graph nodes and the node edge features; the parsing graph node text features and the node edge text features are obtained according to the parsing graph nodes and the node edge feature vectors; the text features cannot be directly involved in the calculation; the parsing graph is encoded by the graph machine learning model, the parsing graph node text features and the node edge text features are encoded respectively, the node text features and the node edge text features description format are unified, the text vector participates in the calculation, and the node text feature vector and the node edge text feature vector are obtained; By combining the bottom-up attention mechanism with the top-down attention mechanism, a combined attention mechanism model is constructed; through the combined attention mechanism model, the key focus area is selected, and the key focus sub-parse graph and effective structure graph information are extracted, including: By combining the bottom-up attention mechanism with the top-down attention mechanism, a combined attention mechanism model is constructed, including: comparing the difference between the front and back frames of the parse graph sequence to obtain the front and back frame difference sequence; selecting the largest difference part in the front and back frame difference sequence as the focus part of the parse graph; combining the bottom-up attention mechanism with the top-down attention mechanism, according to the focus part of the parse graph, using a large language model to determine the key part with the highest correlation with the value dimension in the task scenario, thereby determining the key area from top to bottom; fusing the focus part with the key area to obtain the focus key area; constructing a combined attention mechanism model; through the combined attention mechanism model, extracting sub-parse graphs according to the focus key area to obtain the focus key sub-parse graph; selecting effective structure graph information according to the focus key sub-parse graph.

[0038] In one embodiment, the stable characteristic flow state learning subsystem includes: The hybrid multi-head node extraction module extracts local nodes from the effective structure graph information through the hybrid multi-head attention model according to the effective structure graph information; The sub-dimension weight calculation module obtains the sub-dimension value through the interactive neural network model according to the local node; and obtains the sub-dimension weight through the gate neural network; The analysis graph value calculation module intelligently learns stable dynamic characteristic flow states from the observed data, performs weighted summation of sub-dimension values ​​and sub-dimension weights, obtains the value function value, and obtains the value of the sub-analysis graph.

[0039] The principles and effects of the above technical solution are: a stable feature flow state learning subsystem, including: a hybrid multi-head node extraction module, which extracts local nodes in the effective structure graph information through a hybrid multi-head attention model according to the effective structure graph information; a sub-dimension weight calculation module, which obtains sub-dimension values ​​through an interactive neural network model according to local nodes; obtains sub-dimension weights through a gate neural network; a parsing graph value calculation module, which intelligently learns stable dynamic feature flow states from observed data, performs weighted summation of sub-dimension values ​​and sub-dimension weights, obtains value function values, and obtains the value of the sub-parsing graph.

[0040] In one embodiment, the partial order comparison value prediction subsystem includes: The contrastive learning objective function model module establishes the contrastive learning objective function model by comparing the partial order data according to the value of the sub-analysis graph; The value model parameter calculation module minimizes the contrast learning objective function model and calculates the value model parameters; The multi-level value prediction module builds a multi-level value prediction function model based on the value model parameters; based on the multi-level value prediction function model, it conducts multi-level value dimension prediction to achieve value-driven autonomous intelligent prediction; According to the value model parameters, a multi-level value prediction function model is constructed; according to the multi-level value prediction function model, a multi-level value dimension prediction is performed to realize value-driven autonomous intelligent prediction, including: extracting the analysis graphs of two moments from the analysis graph sequence to obtain the analysis graph of the first moment and the analysis graph of the second moment; calculating the value of the first sub-analysis graph of the analysis graph at the first moment and the value of the second sub-analysis graph; calculating the value difference between the value of the first sub-analysis graph and the value of the second sub-analysis graph, and adding an adaptive factor of the interval between two moments to establish a comparative learning objective function model; minimizing the comparative learning objective function model and adding a mean sparse penalty term; the mean sparse penalty term is the product of the sparse factor and the weight mean; forming a multi-level value prediction function model; according to the multi-level value prediction function model, a multi-level value dimension prediction of the user group is performed to realize value-driven autonomous intelligent prediction.

[0041] The principle and effect of the above technical solution are as follows: the partial order comparison value prediction subsystem includes: a comparison learning objective function model module, which establishes a comparison learning objective function model according to the value of the sub-analysis graph through the comparison of partial order data; a value model parameter operation module, which minimizes the comparison learning objective function model and calculates the value model parameters; a multi-level value prediction module, which constructs a multi-level value prediction function model according to the value model parameters; and performs multi-level value dimension prediction according to the multi-level value prediction function model to realize value-driven autonomous intelligent prediction; a multi-level value prediction function model is constructed according to the value model parameters; and a multi-level value prediction function model is performed according to the multi-level value prediction function model to realize value-driven autonomous intelligent prediction. The current value-driven autonomous intelligent prediction includes: extracting analysis graphs at two moments from the analysis graph sequence to obtain the analysis graph at the first moment and the analysis graph at the second moment; calculating the value of the first sub-analysis graph of the analysis graph at the first moment and the value of the second sub-analysis graph; calculating the value difference between the first sub-analysis graph and the second sub-analysis graph, and adding an adaptive factor for the interval between the two moments to establish a comparative learning objective function model; minimizing the comparative learning objective function model and adding a mean sparse penalty term; the mean sparse penalty term is the product of the sparse factor and the weighted mean; forming a multi-level value prediction function model; based on the multi-level value prediction function model, performing multi-level value dimension prediction of user groups to achieve value-driven autonomous intelligent prediction.

[0042] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and the implementation methods. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and the illustrations shown and described herein.

Claims

1. A general artificial intelligence value function learning prediction method, characterized in that: include: S100, collects user personality characteristic data and obtains observation data through a multi-modal multi-user parallel cloud service platform; S200, uses a graph machine learning model to parse graph encoding and extracts effective structural graph information through a combined attention mechanism model combining bottom-up and top-down methods; S300, based on the effective structure graph information, intelligently learns the stable characteristic flow state from the observation data and obtains the value of the sub-analysis graph; S400, according to the value of the sub-analysis graph, establishes a comparative learning objective function model through the comparison of partial order data; minimizes the comparative learning objective function model and constructs a multi-level value prediction function model; performs multi-level value dimension prediction to realize value-driven autonomous intelligent prediction.

2. A general artificial intelligence value function learning prediction method according to claim 1, characterized in that: S100 includes: S101, setting a multi-modal multi-user parallel cloud service platform; a user accesses the multi-modal multi-user parallel cloud service platform through a network task interface; S102, displaying task scenarios in parallel through multiple network task interfaces; and providing task scenario collection prompts, prompting to set task operations, and obtaining task scenario sequences; S103, extracting a parsing graph sequence corresponding to the task scenario sequence.

3. A general artificial intelligence value function learning prediction method according to claim 1, characterized in that: S200 includes: S201, based on the parsing graph nodes and inter-node edge features represented by text, the parsing graph is encoded by a graph machine learning model to obtain a node text feature vector and an inter-node edge text feature vector; S202, constructing a combined attention mechanism model by combining the bottom-up attention mechanism with the top-down attention mechanism; selecting the key areas of attention through the combined attention mechanism model, and extracting the key sub-parse graph and effective structure graph information.

4. A general artificial intelligence value function learning prediction method according to claim 1, characterized in that: S300 includes: S301, extracting local nodes in the effective structure graph information through a hybrid multi-head attention model according to the effective structure graph information; S302, obtaining sub-dimension values ​​through a neural network model according to local nodes; obtaining sub-dimension weights through a gate neural network; S303, intelligently learn the stable characteristic flow state from the observed data, perform weighted summation of the sub-dimension value and the sub-dimension weight, obtain the value of the value function, and obtain the value of the sub-analysis graph.

5. A general artificial intelligence value function learning prediction method according to claim 1, characterized in that: S400 includes: S401, establishing a comparative learning objective function model by comparing the partial order data according to the value of the sub-analysis graph; S402, minimizing the contrast learning objective function model and calculating the value model parameters; S403, constructing a multi-level value prediction function model according to the value model parameters; performing multi-level value dimension prediction according to the multi-level value prediction function model to achieve value-driven autonomous intelligent prediction; According to the value model parameters, a multi-level value prediction function model is constructed; according to the multi-level value prediction function model, a multi-level value dimension prediction is performed to realize value-driven autonomous intelligent prediction, including: extracting the analysis graphs of two moments from the analysis graph sequence to obtain the analysis graph of the first moment and the analysis graph of the second moment; calculating the value of the first sub-analysis graph of the analysis graph at the first moment and the value of the second sub-analysis graph; calculating the value difference between the value of the first sub-analysis graph and the value of the second sub-analysis graph, and adding an adaptive factor of the interval between two moments to establish a comparative learning objective function model; minimizing the comparative learning objective function model and adding a sparse penalty term; the sparse penalty term is the product of the sparse factor and the weight; forming a multi-level value prediction function model; according to the multi-level value prediction function model, a multi-level value dimension prediction of the user is performed to realize value-driven autonomous intelligent prediction.

6. A general artificial intelligence value function learning prediction system, characterized in that: include: The observation data feature cloud collection subsystem collects user personality feature data and obtains observation data through a multi-modal multi-user parallel cloud service platform; The graph model encoding extraction subsystem uses a graph machine learning model to parse graph encoding and extracts effective structural graph information through a combined attention mechanism model that combines bottom-up and top-down methods. The stable characteristic flow state learning subsystem intelligently learns the stable characteristic flow state from the observed data based on the effective structure diagram information and obtains the value of the sub-analysis diagram; The partial order comparative value prediction subsystem establishes a comparative learning objective function model based on the value of the sub-analysis graph and by comparing the partial order data; minimizes the comparative learning objective function model and constructs a multi-level value prediction function model; performs multi-level value dimension prediction and realizes value-driven autonomous intelligent prediction.

7. A general artificial intelligence value function learning prediction system according to claim 6, characterized in that: Observation data feature cloud collection subsystem, including: A multi-modal multi-user parallel cloud interface module is used to set up a multi-modal multi-user parallel cloud service platform; users access the multi-modal multi-user parallel cloud service platform through a network task interface; The task interface parallel display module displays task scenarios in parallel through multiple network task interfaces; and provides task scenario collection prompts, prompts to set task operations, and obtains task scenario sequences; The scene parsing graph sequence extraction module extracts the parsing graph sequence corresponding to the task scene sequence.

8. A general artificial intelligence value function learning prediction system according to claim 6, characterized in that: Stable characteristic flow state learning subsystem, including: The parsing graph encoding module encodes the parsing graph nodes and inter-node edge features based on the text representation, and obtains the node text feature vector and the inter-node edge text feature vector through the graph machine learning model; The combined attention mechanism extraction module constructs a combined attention mechanism model by combining the bottom-up attention mechanism with the top-down attention mechanism. Through the combined attention mechanism model, the key areas of attention are selected, and the key sub-parse graphs and effective structure graph information are extracted.

9. A general artificial intelligence value function learning prediction system according to claim 6, characterized in that: Stable characteristic flow state learning subsystem, including: The hybrid multi-head node extraction module extracts local nodes from the effective structure graph information through the hybrid multi-head attention model according to the effective structure graph information; The sub-dimension weight calculation module obtains the sub-dimension value through the interactive neural network model according to the local node; and obtains the sub-dimension weight through the gate neural network; The analysis graph value calculation module intelligently learns the stable characteristic flow state from the observed data, performs weighted summation of the sub-dimension value and the sub-dimension weight, obtains the value function value, and obtains the value of the sub-analysis graph.

10. A general artificial intelligence value function learning prediction system according to claim 6, characterized in that: Partial order comparison value prediction subsystem, including: The contrastive learning objective function model module establishes the contrastive learning objective function model by comparing the partial order data according to the value of the sub-analysis graph; The value model parameter calculation module minimizes the contrast learning objective function model and calculates the value model parameters; The multi-level value prediction module builds a multi-level value prediction function model based on the value model parameters; based on the multi-level value prediction function model, it conducts multi-level value dimension prediction to achieve value-driven autonomous intelligent prediction; According to the value model parameters, a multi-level value prediction function model is constructed; according to the multi-level value prediction function model, a multi-level value dimension prediction is performed to realize value-driven autonomous intelligent prediction, including: extracting the analysis graphs of two moments from the analysis graph sequence to obtain the analysis graph of the first moment and the analysis graph of the second moment; calculating the value of the first sub-analysis graph of the analysis graph at the first moment and the value of the second sub-analysis graph; calculating the value difference between the value of the first sub-analysis graph and the value of the second sub-analysis graph, and adding an adaptive factor of the interval between two moments to establish a comparative learning objective function model; minimizing the comparative learning objective function model and adding a sparse penalty term; the sparse penalty term is the product of the sparse factor and the weight; forming a multi-level value prediction function model; according to the multi-level value prediction function model, a multi-level value dimension prediction of the user group is performed to realize value-driven autonomous intelligent prediction.

Images