A value-driven autonomous intelligent system

Abstract

This invention provides a value-driven autonomous intelligent system, comprising: an intelligent agent, wherein the intelligent agent includes: a value extraction module, a backsampling generation module, and a planning module; wherein the value extraction module extracts value from the scene to obtain predicted value; the backsampling generation module is used to backsample and generate a target state analytical graph based on given expected value and predicted value; the planning module is used to plan and generate action trajectories based on the initial analytical graph and the target state analytical graph. This value-driven autonomous intelligent system achieves more intelligent communication by driving the intelligent agent's behavior through value.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology for achieving autonomous intelligence, and particularly to a value-driven autonomous intelligence system. Background Technology

[0002] Research on autonomous intelligence can be divided into two stages. The first stage focuses on intelligent systems centered on cognitive architectures, which posits that implementing several cognitive modules is sufficient for system autonomy. However, these modules lack a stable intrinsic driver module to guide the system in proposing new tasks. The second stage focuses on agent systems centered on large language models. These models instruct agents on which tasks to perform in different contexts, enabling autonomous communication. However, the black-box nature of large language models and their lack of interpretability make it unclear where the autonomy originates. The following is a brief analysis of typical works from these two stages.

[0003] Cognitive Architecture Stage: Since the birth of artificial intelligence, more than 80 cognitive architecture models have been proposed. Most of these models remain at the conceptual level, either being validated on very simple examples or remaining at the level of thought (for details, please refer to "A Review of 40 Years of Cognitive Architecture Research"). Some well-known cognitive architectures include: ACT-R (Adaptive Control of Thought-Rational), developed by Carnegie Mellon University, used to simulate human thought and behavior processes, including modules such as perception, memory, and decision-making; SOAR, a general problem-solving architecture developed by the University of Michigan, emphasizing symbolic operations and suitable for complex task modeling; LIDA (Learning Intelligent Distribution Agent), based on the global workspace theory in cognitive science, suitable for multi-task scenarios; CLARION (Connectionist Learning with Adaptive Rule Induction ON-line), combining symbolic and connectionist approaches to simulate human conscious and unconscious processes; and SPAUN (Semantic Pointer Architecture Unified Network), developed by the University of Buffalo, which simulates the cognitive functions of the human brain, combining neural network structures and biological constraints.

[0004] Large Language Model Stage: With the recent rise of large language models, cognitive architectures based on these models, incorporating modules for memory, reasoning, and reflection, have been proposed. These are largely reimplementations of traditional cognitive architectures; see "Cognitive Architectures for Language Agents" for details. A representative work on autonomous agents based on large language models is Stanford's AI Village, launched in 2023. Researchers provided each agent with a brief biography, including name, age, occupation, family, interests, and some habits, allowing them to develop their own. These agents then relied on a large language model to generate their behaviors based on their prescribed biographies, simulating human-like actions: waking up, making breakfast, going to work, eating lunch, and chatting with other agents they encountered. They also remembered events, reflected on them, and made plans. For example, when the researcher in charge of the landscape suggested a character plan a Valentine's Day party, she invited friends and acquaintances, many of whom showed up at the right time and place. Following AI Village, a number of similar autonomous agents driven by large language models emerged.

[0005] Currently, autonomous intelligent agents based on large language models use prompt word technology, which involves manually writing specific behavioral steps to drive the agent's behavior. However, when applied to communication-type intelligent systems, this approach fails to achieve more intelligent communication. Summary of the Invention

[0006] One of the objectives of this invention is to provide a value-driven autonomous intelligent system that enables more intelligent communication by driving the behavior of intelligent agents through value.

[0007] This invention provides a value-driven autonomous intelligent system, comprising: an intelligent agent, wherein the intelligent agent includes: a value extraction module, a backsampling generation module, and a planning module;

[0008] The value extraction module extracts value from the scene to obtain the predicted value; the backsampling generation module is used to backsample and generate the target state analysis map based on the given expected value and predicted value; and the planning module is used to plan and generate the action trajectory based on the initial analysis map and the target state analysis map.

[0009] Preferably, the value extraction module extracts value from the scenario to obtain predicted value, and performs the following operations:

[0010] The scene is analyzed to generate a large analytical graph;

[0011] The value of each region in the scene is determined based on the large analytical graph;

[0012] Extract the value of the region of interest from the large analytical map as the predictive value.

[0013] Preferably, the backsampling generation module is used to backsample and generate a target state analytical map based on a given expected value and predicted value, including:

[0014] Determine the difference between expected value and predicted value;

[0015] The probability distribution of the target state analytical diagram is established based on the differences;

[0016] Based on the probability distribution of the target state analytical graph, the sub-analytical graphs corresponding to the large analytical graph are sampled to obtain the target state analytical graph.

[0017] Preferred, pre-configured algorithm rules include: genetic algorithm or Monte Carlo algorithm.

[0018] Preferably, the probability distribution is expanded to obtain a functional form before calculation.

[0019] Preferably, the intelligent agent further includes: an iteration module; the iteration module is used to update the initial state and target state after the action trajectory is performed, and to cyclically control the actions of the value extraction module, the back sampling generation module and the control planning module; the iteration ends when the difference between the expected value and the predicted value is within a preset range.

[0020] Preferred, value-driven autonomous intelligent systems also include:

[0021] The system comprises an input module, an agent construction module, and an output module; the agent construction module includes a value calibration unit, a value extraction unit, and a construction unit.

[0022] The value calibration unit is used to generate calibrated values; the value extraction unit is used to retrieve relevant values ​​from the value library based on the calibrated values; and the construction unit is used to construct an intelligent agent based on the calibrated values ​​and / or the retrieved values.

[0023] Preferably, the steps for constructing value in the value pool are as follows:

[0024] Group the data samples to obtain multiple sample groups;

[0025] Each sample group corresponds to a training value.

[0026] The preferred, valuable training steps are as follows:

[0027] Extract flow patterns from data samples;

[0028] Extracting value scalars from the flow state;

[0029] The pre-configured model is trained based on the value scalar.

[0030] Preferably, the value extraction unit is also used to extract values ​​marked with representative labels from the value library; the value calibration unit generates multiple calibrated values; when multiple values ​​are generated, each value is output and weight configuration is received.

[0031] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0032] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0033] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0034] Figure 1 This is a schematic diagram of a value-driven autonomous intelligent system according to an embodiment of the present invention;

[0035] Figure 2 This is a driving principle diagram based on a specific scenario in an embodiment of the present invention;

[0036] Figure 3 This is the target analytical graph sampling process based on value differences in this embodiment of the invention;

[0037] Figure 4 This is a schematic diagram illustrating the autonomous behavior generated based on iterative optimization of target value in an embodiment of the present invention;

[0038] Figure 5 This is a schematic diagram of yet another value-driven autonomous intelligent system in an embodiment of the present invention;

[0039] Figure 6 This is a diagram illustrating the training steps for value in an embodiment of the present invention. Detailed Implementation

[0040] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0041] This invention provides a value-driven autonomous intelligent system, such as... Figure 1 As shown, it includes: an intelligent agent, wherein the intelligent agent includes: a value extraction module 1, a backsampling generation module 2, and a planning module 3;

[0042] Among them, the value extraction module 1 extracts value from the scene to obtain the predicted value; the backsampling generation module 2 is used to backsample and generate the target state analysis map based on the given expected value and predicted value; and the planning module 3 is used to plan and generate the action trajectory based on the initial analysis map and the target state analysis map.

[0043] The value-driven autonomous intelligent system of this invention drives the actions of an intelligent agent through value. The principle is as follows: First, determine the difference between the value contained in the scene and the given expected value. Based on this difference, backsample to generate a target state analytical map. Then, the planning module plans and generates an action trajectory based on the initial analytical map and the target state analytical map. That is, given the expected value V*, predict the value V, and the value difference delta_V=V*-V, combined with the geometric personality trait vector h, a target state analytical map (Goal PG) can be generated through backsampled sampling. Given the initial analytical map LPG and the target analytical map GoalPG, an action trajectory is planned and generated to reach the target state from the initial state. (See...) Figure 2 Furthermore, the figure also shows that from the initial parsing graph (LPG), a sub-parsing graph (eLPG) is obtained through encoder processing. Then, a top-down or bottom-up attention mechanism is introduced to obtain a sub-parsing graph with an attention mechanism. From the sub-parsing graph, a learning flow state is obtained. The learning flow state and the geometric personality trait vector h are combined and processed through a value function MLP to obtain the predicted value. The difference between the expected value and the predicted value is given as follows: Then, by inversely deducing the differences, the selected sub-PG (the smallest unit constituting the analytical graph) is derived, and sampling it yields the target analytical graph, Goal PG. The planning module generates action trajectories based on causal chain search for task planning: given an initial state and a target state, to establish the transition trajectory between these two states, a causal chain needs to be searched in the causal database to establish the connection between the two states. A causal chain is a very stable state transition relationship, and using a causal chain greatly reduces the search space; the given initial state is represented as the initial analytical graph, and the target state as the target analytical graph. To implement the causal chain, a pre-configured causal database can be used; that is, given an initial state and a target state, to establish the transition trajectory between these two states, a causal chain needs to be searched in the causal database to establish the connection between the two states. Furthermore, to reduce the sampling space, sampling can be performed only in the region of interest, and using the initial state as the initial sample will greatly accelerate the sampling process. In this context, a scene can be specifically represented as one or more images. When an autonomous intelligent system is applied in a virtual communication environment, the scene is acquired through user input interacting with the agent or automatically generated within the virtual environment. When an autonomous intelligent system is applied in a real-world communication scenario (installed in an intelligent robot), the scene is acquired through a correspondingly configured image acquisition module. For example, in a virtual environment, the agent can complete a series of tasks. A specific task is for the agent to tidy up items on a table. It first learns the value dimension function of "tidyness," then samples various forms of tidyness based on this function, including placing items in a line or a circle. Based on this line or circle, it plans specific operational steps and executes the task, demonstrating value-driven autonomy.

[0044] The value extraction module extracts value from the scenario to obtain predicted value, and performs the following operations:

[0045] The scene is analyzed to generate a large analytical graph;

[0046] The value of each region in the scene is determined based on the large analytical graph;

[0047] Extract the value of the region of interest from the large analytical map as the predictive value.

[0048] Extracting the large analytical graph (LPG) of the scene depicts the value of each region in the scene, thus forming a value map. When the value Vi of the region of interest differs from the expected value V*, an expected difference (or expected gap) delta_V is generated. This expected difference will drive the agent to generate the target task state, thereby generating a new task.

[0049] The backsampling generation module is used to backsample and generate a target state analytical graph based on the given expected value and predicted value, including:

[0050] Determine the difference between expected value and predicted value;

[0051] The probability distribution of the target state analytical diagram is established based on the differences;

[0052] Based on the probability distribution of the target state analytical graph, the sub-analytical graphs corresponding to the large analytical graph are sampled to obtain the target state analytical graph.

[0053] Pre-configured algorithm rules include: genetic algorithm or Monte Carlo algorithm.

[0054] Before calculating the probability distribution, the probability distribution is expanded to obtain its functional form.

[0055] Based on delta_V, construct the probability distribution p(pg*) of the target analytical graph pg*, such as Figure 3 As shown, sampling is performed using MCMC (Markov Chain Monte Carlo method). ; indicates ; One point to note is... Sampling is guided by scene / task syntax to ensure effectiveness. Focus on children Sampling instead of the entire LPG Conditional sampling is performed based on existing LPGs to simplify subsequent task planning. The distribution function can be further expanded to obtain a function of pg* as a variable. Here, a genetic algorithm or the Monte Carlo algorithm can be used to calculate the pg* with the maximum probability. Taking the Monte Carlo algorithm as an example, three aspects need to be considered during the sampling process. First, the sampling process is guided by a task grammar rule, which ensures the rationality of the sampling task. Second, the entire sampling process only samples the sub-analytical graph of interest, without sampling the entire large analytical graph. Finally, sampling based on the analytical graph in its initial state is more efficient.

[0056] In one embodiment, the intelligent agent further includes: an iteration module; the iteration module is used to update the initial state and target state after the action trajectory is performed, and to cyclically control the actions of the value extraction module, the backsampling generation module and the control planning module; the iteration ends when the difference between the expected value and the predicted value is within a preset range.

[0057] See Figure 4 For searching from arrive To find effective causal link paths, a forward / backward search strategy is employed, for example... , RRT, CEM; during the search, evaluate ,calculate If the condition is met, stop; otherwise, proceed based on the newly generated... Return to sampling new Since the sampled target task state may not fully reflect the target value, and the target task state cannot be perfectly realized in task planning, the uncertainty brought about by these two factors means that the value-driven task cannot be completed in one go. It is necessary to iteratively update the initial state and the target state, and gradually approach the target value, thereby implementing multiple tasks. When the distance between the actual value and the target value is within an acceptable range, the task generation ends.

[0058] In one embodiment, a value-driven autonomous intelligent system, such as Figure 5 As shown, it also includes: an input module 11, an agent construction module 12, and an output module 13; wherein, the agent construction module 12 includes: a value calibration unit 121, a value extraction unit 122, and a construction unit 123;

[0059] Among them, the value calibration unit 121 is used to generate a calibrated value; the value extraction unit 122 is used to retrieve relevant values ​​from the value library based on the calibrated value; and the construction unit 123 is used to construct an intelligent agent based on the calibrated value and / or the retrieved value.

[0060] The value-driven autonomous intelligent system in this embodiment retrieves relevant values ​​from a pre-configured value library through a value extraction unit based on human-defined values, thereby constructing multiple intelligent agents and achieving more intelligent communication. Through multiple intelligent agents guided by different values, it realizes multifaceted insights into the same problem, thus achieving more intelligent communication.

[0061] The working principle of the intelligent agent can be understood as follows: First, a large parsing graph (LPG) of the given initial scene is extracted. This graph depicts the value of each region in the scene, thus forming a value map. When the value Vi of the region of interest differs from the expected value V*, an expected difference (or expected gap) delta_V is generated. This expected difference drives the agent to generate a target task state, thereby generating a new task. That is, given the expected value V*, the agent predicts the value V, and the value difference delta_V = V* - V. The geometric personality trait vector h can be backsampled to generate the target state parsing graph. Given the initial parsing graph LPG and the target parsing graph Goal PG, an action trajectory is planned and generated to reach the target state from the initial state. Among them, the probability distribution p(pg*) of the target parsing graph pg* is established based on delta_V, such as... Figure 3 As shown, the form of this distribution function can be further expanded to obtain a function of pg* as a variable. Here, a genetic algorithm or the Monte Carlo algorithm can be used to calculate the pg* with the maximum probability. Taking the Monte Carlo algorithm as an example, three aspects need to be considered during the sampling process. First, the sampling process is carried out under the guidance of a task syntax rule, which ensures the rationality of the sampling task. Second, the entire sampling process only samples the sub-analytic graph of interest, without sampling the entire large analytic graph. Finally, sampling based on the analytic graph of the initial state is more efficient. Since the sampled target task state may not fully reflect the target value, and the target task state cannot be perfectly realized in task planning, the uncertainty brought about by these two factors means that the value-driven task cannot be completed in one go. It is necessary to iteratively update the initial state and the target state, step by step approaching the target value, thereby implementing multiple tasks. When the distance between the actual value and the target value is within an acceptable range, the task generation ends.

[0062] In one embodiment, the steps for constructing value in the value pool are as follows:

[0063] Group the data samples to obtain multiple sample groups;

[0064] Each sample group corresponds to a training value.

[0065] Data samples can be grouped randomly. That is, you can set how many groups of samples are needed and then randomly select samples to fill them. Each sample group can be trained to have a value, which can then be stored in the value library for later use.

[0066] In one embodiment, such as Figure 6 As shown, the training steps for value are as follows:

[0067] Step 1: Extract the flow regime from the data sample;

[0068] Step 2: Extracting value scalars from the flow regime;

[0069] Step 3: Train the pre-configured model based on the value scalar.

[0070] The specific process of generating value, or the value function, is as follows: Data Sample: This is the raw data collected from the environment or task, which may be images, audio, video, text, or structured data. Fluent / Latent Variable: These are the underlying dynamic features or hidden patterns behind the data, reflecting changes in the data under different states. Latent variables represent a certain latent attribute or state of a data sample. Value Scalar: This is a numerical value representing the value of a data sample, usually obtained through the model's output, representing the quality of the target state or behavior corresponding to that sample. The steps for generating latent variables and sample data are as follows: For sequence data (e.g., time series, text), we use RNN or Transformer, where latent variables serve as input to the network's hidden layers, capturing dependencies and state transitions in the time series. For image data, we use CNN or UNet, where latent variables serve as additional feature encoding, helping to extract key features from the image and generate data samples. For graph data (such as social networks or transportation networks), GNNs (Graph Neural Networks) can be used to process the relationships between nodes in the data through the graph structure, thereby inferring latent variables and sample data. These different architectures can generate latent variables that fit different modalities of data, helping to establish accurate value functions.

[0071] In one embodiment, the value calibration unit is used to generate a calibrated value by performing the following operations:

[0072] Output the pre-defined value survey questionnaire items;

[0073] Receive feedback information for each item in the value survey questionnaire;

[0074] Based on feedback information and a pre-configured parameter mapping table, determine the parameter data;

[0075] Based on the defined parameter data, a calibrated value is generated.

[0076] The calibrated value is essentially user-oriented and constructed based on the information provided by the user. To understand user information, value survey questionnaire items can be configured. Users configure different values ​​based on the value survey questionnaire items. There are multiple value survey questionnaire items, and each value survey questionnaire item has multiple corresponding feedback items, each with different parameter data values. Then, by combining all the survey questionnaire items, the calibrated value can be obtained. Furthermore, users can select scenarios in the early stage and provide different value survey questionnaire items based on different scenarios.

[0077] To extract relevant value from the value repository, the value extraction unit retrieves relevant value from the value repository based on the calibrated value, performing the following operations:

[0078] The parameter data of the calibrated value are arranged in a pre-configured order to form the first dataset;

[0079] Arrange the parameter data of each value in the value library in a pre-configured order to form a second dataset;

[0080] Calculate the similarity between the first dataset and the second dataset;

[0081] Extract the value that minimizes the difference between various relevance thresholds specified in a pre-configured relevance set.

[0082] By quantifying each value to form a first dataset and a second dataset, the values ​​in the value library are compared with the labeled values ​​using similarity calculation. In addition, when retrieving relevant values, a relevant set is used as a guide. Generally, relevance is defined as the highest similarity; relevance can also be expressed as the highest difference, i.e., the lowest similarity, which is optimally zero. Therefore, this embodiment uses a relevant set to determine the relevance, for example, configuring 0%, 20%, 50%, and 100% in the relevant set. This requires retrieving four values, namely the ones with the highest similarity, the lowest similarity, and those close to 20% and 50%. Furthermore, the similarity can be calculated using the cosine similarity method.

[0083] In one embodiment, sample data can be collected through a solicitation process, with the following steps: Contextual data is released to solicit social interaction data. Data on big data platforms is readily available and readily accessible; however, data obtained through contextual data release is slower. This invention combines both collection methods, with the following steps: Data from the corresponding category on the big data platform is extracted, and statistical parameters are determined. When the statistical parameters do not meet preset conditions, the contextual data is released to collect the solicited data. The statistical parameters are updated based on the collected data, and the solicitation ends when the preset conditions are met. The statistical parameters include one or more combinations of: total data count, the number of data points for each different interactive decision-making data set, and the number of interactive decision-making data sets. The preset conditions include one or more combinations of: a data threshold corresponding to the total data count, a data threshold corresponding to each different interactive decision-making data set, and a quantity threshold corresponding to the number of interactive decision-making data sets.

[0084] Furthermore, the validity of the data needs to be verified during data collection. A specific verification method can be used to indirectly verify the data by analyzing the users who provide it. The specific analysis process is as follows: Obtain the user's permission information; based on the permission information and a pre-configured first scoring table, determine the first score value; obtain the user's data provided in the collection module corresponding to historical collections, calculate the similarity between the historical data and the currently provided data, and determine the second score value based on the similarity and a pre-configured second scoring table; construct an analysis vector based on the time interval between the user's data provision in the collection module (pre-set number of times), and then use the analysis vector to select from a pre-configured third scoring database. The system retrieves the corresponding third score value. Verification is successful when the sum of the first, second, and third scores exceeds a preset scoring threshold. In the second score table, each second score corresponds one-to-one with the similarity score. A specific intermediate value can be configured (e.g., a similarity of 0.8 corresponds to a score of 100). Starting from this intermediate similarity value, the second score gradually decreases as the similarity increases or decreases. In the third score library, each third score corresponds one-to-one with the analysis vector. This third score library is pre-configured. The configuration follows these rules: the shorter the average interval, the lower the third score; for the same average time interval, the smaller the interval, the lower the third score.

[0085] User permission information is dynamically determined based on initial configuration values ​​and user behavior analysis to ensure accuracy. The specific permission value is determined by counting the number of times user-provided data is judged as accurate and the number of times it is judged as inaccurate. When the number of inaccurate judgments exceeds a preset threshold, the permission value is set to zero; otherwise, it is the sum of the initial configuration value, the number of accurate judgments multiplied by a first adjustment value, and the number of accurate judgments multiplied by a second adjustment value. The first adjustment value is positive, the second adjustment value is negative, and the absolute value of the first adjustment value is less than the absolute value of the second adjustment value. By using the difference between the first and second adjustment values, a punitive reduction in permission value can be applied to malicious data provision to prevent future malicious data provision.

[0086] In practical use, different value functions are invoked according to different current scenarios to make targeted decisions and improve the level of humanization; and different scenario configurations correspond to different value functions for different personalities to adapt to the interactive experience of different personalities.

[0087] In one embodiment, the value extraction unit is also used to extract the value marked with representative labels from the value library.

[0088] By labeling representative tags and outputting them synchronously, we can achieve feedback on representative values; the labeling of representative tags is done by professionals after professional analysis.

[0089] To enable communication between the autonomous intelligent system and the user, in one embodiment, the input module is used to receive input data from the user; the output module is used to output the output data of each intelligent agent in response to the input data.

[0090] The autonomous intelligent system realizes input and output through input and output data. The input module transmits the input data to each intelligent agent respectively; the output module receives the output of each intelligent agent.

[0091] In one embodiment, the value calibration unit generates multiple calibrated values; when multiple values ​​are generated, each value is output and a weight configuration is received.

[0092] In reality, people are complex, and an individual may possess multiple thoughts, usually with one dominant one. Therefore, the value assessment unit's questionnaire items can select multiple feedback items, generating multiple assessed values. These values ​​are then described and output, with users assigning weights, the sum of which is 1. During this process, a pre-configured value-description mapping library is used to retrieve the corresponding value description from the library based on the generated assessed values. Because the value function itself is data, not everyone can understand it. Describing and illustrating each data point in the value function provides users with an intuitive understanding, facilitating weight allocation. For example, in the value function, data A represents the choice at a fork in the road, with 0 representing the right and 1 representing the left. The corresponding descriptions are: when encountering a fork in the road, the bias is towards the right; when encountering a fork in the road, the bias is towards the left. Configuring the bias weights for multiple values ​​allows for further consideration of the randomness of the choice.

[0093] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A value-driven autonomous intelligent system, characterized in that, include: The intelligent agent includes: a value extraction module, a backsampling generation module, and a planning module; The value extraction module extracts value from the scene to obtain the predicted value; the scene is represented by one or more images; the backsampling generation module is used to backsample and generate a target state analytical map based on the given expected value and predicted value; the planning module is used to plan and generate action trajectories based on the initial analytical map and the target state analytical map. Also includes: The system comprises an input module, an agent construction module, and an output module; the agent construction module includes a value calibration unit, a value extraction unit, and a construction unit. The value calibration unit is used to generate a calibrated value; the value extraction unit is used to retrieve relevant values ​​from the value library based on the calibrated value; and the construction unit is used to construct an intelligent agent based on the calibrated value and / or the retrieved value. The steps for constructing value in the value pool are as follows: Group the data samples to obtain multiple sample groups; Each sample group corresponds to a training value; The steps for obtaining data samples are as follows: extract data from the corresponding category of the big data platform and determine the statistical parameters of the data; when the statistical parameters do not meet the preset conditions, publish the scenario data and collect the collected data; update the statistical parameters based on the collected data, and end the collection when the preset conditions are met; wherein, the statistical parameters include: one or more of the following combinations: total number of data, number of data for each different interactive decision data, and number of interactive decision data; the preset conditions include one or more of the following combinations: a data threshold corresponding to the total number of data, a data threshold corresponding to each different interactive decision data, and a number threshold corresponding to the number of interactive decision data.

2. The value-driven autonomous intelligent system as described in claim 1, characterized in that, The value extraction module extracts value from the scenario to obtain predicted value, and performs the following operations: The scene is analyzed to generate a large analytical graph; The value of each region in the scene is determined based on the large analytical graph; Extract the value of the region of interest from the large analytical map as the predictive value.

3. The value-driven autonomous intelligent system as described in claim 2, characterized in that, The backsampling generation module is used to backsample and generate a target state analytical graph based on given expected and predicted values, including: Determine the difference between expected value and predicted value; The probability distribution of the target state analytical diagram is established based on the differences; Based on the probability distribution of the target state analytical graph, the sub-analytical graphs corresponding to the large analytical graph are sampled to obtain the target state analytical graph.

4. The value-driven autonomous intelligent system as described in claim 3, characterized in that, Before calculating the probability distribution, the probability distribution is expanded to obtain its functional form.

5. The value-driven autonomous intelligent system as described in claim 1, characterized in that, The intelligent agent also includes: an iteration module; the iteration module is used to update the initial state and target state after the action trajectory, and to cyclically control the actions of the value extraction module, the back sampling generation module and the control planning module; the iteration ends when the difference between the expected value and the predicted value is within a preset range.

6. The value-driven autonomous intelligent system as described in claim 1, characterized in that, The training steps for value are as follows: Extract flow patterns from data samples; Extracting value scalars from the flow state; The pre-configured model is trained based on the value scalar.

7. The value-driven autonomous intelligent system as described in claim 1, characterized in that, The value extraction unit is also used to extract values ​​labeled with representative tags from the value library; the value calibration unit generates multiple calibrated values; when multiple values ​​are generated, each value is output and the weight configuration is received.

Images