Observation method and device of intelligent agent, storage medium and electronic equipment

Abstract

The invention provides an intelligent agent observation method and device, a storage medium and electronic equipment, and belongs to the field of artificial intelligence, and the method comprises the steps: determining sensor information corresponding to an intelligent agent, including sensor data corresponding to a plurality of information dimensions; determining internal state information of the intelligent agent; encoding the sensor information and the internal state information to obtain sensor encoding information and internal state encoding information; determining a decision result; constructing a decision graph based on the sensor coding information, the internal state coding information and the decision result; performing feature extraction on the sensor information to obtain feature information; performing intention inference based on the feature information to obtain an intention inference result; determining decision path information based on the behavior evaluation map; and based on the decision verification rule, performing decision verification on the decision result to obtain a decision verification result. By applying the method provided by the invention, multi-dimensional observation can be carried out on the agent decision on the basis of multi-dimensional data in real time, and the safety and reliability of the system can be improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an observation method and device for an intelligent body, a storage medium, and an electronic device. Background Art

[0002] With the development of artificial intelligence (AI) technology, AI agents have been widely used in many fields to make decisions. AI agents are agents that can perceive their environment and take actions to achieve specific goals. They can also be simply called agents.

[0003] In intelligent agent applications, it's often necessary to observe the agent's behavior to assess whether its decisions meet business requirements. Currently, this observation is typically achieved through log analysis systems, which record and analyze log information during the agent's operation.

[0004] In the existing intelligent agent observation method, the observation of the intelligent agent relies on log information, and the log information is generated after the intelligent agent performs corresponding operations. It usually takes a certain amount of time to obtain the log information used to observe the intelligent agent and complete the log analysis, resulting in a high delay in the observation of the intelligent agent, making it difficult to detect and handle emergencies in a timely manner, resulting in low security and reliability of the system. Summary of the Invention

[0005] In view of this, an embodiment of the present invention provides an observation method for an intelligent agent to solve the problem that the existing intelligent agent observation method relies on log information to observe the intelligent agent, has a high time delay, and leads to low system security and reliability.

[0006] The embodiment of the present invention further provides an observation device for an intelligent agent to ensure the practical implementation and application of the above method.

[0007] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0008] An observation method for an intelligent agent, comprising:

[0009] Determining sensor information corresponding to the intelligent agent to be observed; the sensor information includes sensor data corresponding to multiple information dimensions;

[0010] Determining internal state information corresponding to the agent;

[0011] Encode the sensor information and the internal state information respectively to obtain sensor encoding information corresponding to the sensor information and internal state encoding information corresponding to the internal state information;

[0012] Determining a decision result corresponding to the agent;

[0013] Constructing a decision graph corresponding to the agent based on the sensor encoding information, the internal state encoding information and the decision result;

[0014] Performing feature extraction on the sensor information to obtain feature information corresponding to the sensor information;

[0015] Based on the feature information, performing intention inference on the agent to obtain an intention inference result corresponding to the agent;

[0016] Determine the decision path information corresponding to the decision result based on a pre-built behavior evaluation map;

[0017] Based on a preset decision verification rule, the decision result is verified to obtain a decision verification result corresponding to the intelligent agent;

[0018] The decision map, the intention inference result, the decision path information and the decision verification result are used as the observation results corresponding to the intelligent agent.

[0019] In the above method, optionally, each of the information dimensions includes an image information dimension and a voice information dimension, and extracting features from the sensor information to obtain feature information corresponding to the sensor information includes:

[0020] For the sensor data corresponding to the image information dimension, applying a pre-built image feature extraction model to perform image feature extraction on the sensor data to obtain image features corresponding to the image information dimension; the image feature extraction model is a feature extraction model built based on a convolutional neural network;

[0021] For the sensor data corresponding to the voice information dimension, a pre-built voice feature extraction model is applied to the sensor data to extract voice features, thereby obtaining voice features corresponding to the voice information dimension; the voice feature extraction model is a feature extraction model built based on a long short-term memory network;

[0022] The image features and the voice features are used as feature information corresponding to the agent.

[0023] Optionally, the method described above includes performing intention inference on the agent based on the feature information to obtain an intention inference result corresponding to the agent, including:

[0024] Applying a pre-built intention inference model to perform attention inference processing on the feature information to obtain attention weights corresponding to the features of each region in the feature information; the intention inference model is a model built based on the attention mechanism;

[0025] Compare the attention weights corresponding to the respective regional features, and use the regional feature with the relatively larger corresponding attention weight as the target regional feature;

[0026] Based on the target area characteristics, the decision intention corresponding to the intelligent agent is determined, and the decision intention is used as the intention inference result corresponding to the intelligent agent.

[0027] In the above method, optionally, the process of constructing the behavior evaluation map includes:

[0028] Determining each operating state corresponding to the agent;

[0029] Determine a value data set corresponding to each of the operating states; the value data set corresponding to each of the operating states includes a Q value corresponding to each action in the operating state;

[0030] The behavior evaluation map is constructed based on the value data sets corresponding to each of the operating states.

[0031] Optionally, in the above method, determining the decision path information corresponding to the decision result based on a pre-built behavior evaluation graph includes:

[0032] Determining behavior evaluation information corresponding to the internal state information based on the behavior evaluation map;

[0033] Based on the behavior evaluation information, the decision result is traced back to obtain decision path information corresponding to the decision result.

[0034] The above method may optionally further include:

[0035] Based on the decision verification result, the decision of the intelligent agent is optimized.

[0036] The above method may optionally further include:

[0037] During the process of observing the intelligent agent, various data are transmitted based on a preset lightweight bus.

[0038] An observation device for an intelligent agent, comprising:

[0039] A first determining unit is configured to determine sensor information corresponding to the agent to be observed; the sensor information includes sensor data corresponding to multiple information dimensions;

[0040] A second determining unit, configured to determine internal state information corresponding to the agent;

[0041] an information encoding unit, configured to encode the sensor information and the internal state information respectively, to obtain sensor encoding information corresponding to the sensor information and internal state encoding information corresponding to the internal state information;

[0042] A third determining unit, configured to determine a decision result corresponding to the agent;

[0043] A graph construction unit, configured to construct a decision graph corresponding to the agent based on the sensor encoding information, the internal state encoding information, and the decision result;

[0044] a feature extraction unit, configured to extract features from the sensor information to obtain feature information corresponding to the sensor information;

[0045] an intention inference unit, configured to infer the intention of the agent based on the feature information and obtain an intention inference result corresponding to the agent;

[0046] a fourth determining unit, configured to determine decision path information corresponding to the decision result based on a pre-built behavior evaluation graph;

[0047] A decision verification unit, configured to perform decision verification on the decision result based on a preset decision verification rule to obtain a decision verification result corresponding to the intelligent agent;

[0048] The fifth determination unit is used to use the decision map, the intention inference result, the decision path information and the decision verification result as the observation results corresponding to the intelligent agent.

[0049] A storage medium includes stored instructions, wherein when the instructions are executed, the device where the storage medium is located is controlled to execute the above-mentioned observation method for an intelligent agent.

[0050] An electronic device includes a memory and one or more instructions, wherein the one or more instructions are stored in the memory and configured to be executed by one or more processors to perform the above-mentioned intelligent agent observation method.

[0051] An observation method for an intelligent agent provided based on the above-mentioned embodiment of the present invention includes: determining the sensor information corresponding to the intelligent agent to be observed; the sensor information includes sensor data corresponding to multiple information dimensions; determining the internal state information corresponding to the intelligent agent; encoding the sensor information and the internal state information respectively to obtain sensor coding information corresponding to the sensor information and internal state coding information corresponding to the internal state information; determining the decision result corresponding to the intelligent agent; constructing a decision map corresponding to the intelligent agent based on the sensor coding information, the internal state coding information and the decision result; performing feature extraction on the sensor information to obtain feature information corresponding to the sensor information; based on the feature information, inferring the intention of the intelligent agent to obtain the intention inference result corresponding to the intelligent agent; determining the decision path information corresponding to the decision result based on a pre-constructed behavior evaluation map; performing decision verification on the decision result based on a preset decision verification rule to obtain the decision verification result corresponding to the intelligent agent; and using the decision map, intention inference result, decision path information and decision verification result as the observation result corresponding to the intelligent agent. By applying the method provided in the embodiment of the present invention, a corresponding decision map can be constructed based on sensor data, the internal state information of the intelligent agent and the decision results, the intention of the intelligent agent can be inferred and the decision path can be traced back, and the decision results can be verified at the same time. Finally, an observation result including the decision map, intention inference results, decision path information and decision verification results is obtained. After the intelligent agent makes a decision, multi-dimensional observation of the intelligent agent can be performed in real time, which is conducive to timely discovery and handling of emergencies and improving the safety and reliability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0053] Figure 1 A flow chart of an observation method for an intelligent agent provided by an embodiment of the present invention;

[0054] Figure 2 An example diagram of an observation process of an intelligent agent provided by an embodiment of the present invention;

[0055] Figure 3 An example diagram of a state coding process provided by an embodiment of the present invention;

[0056] Figure 4 A schematic structural diagram of an observation device for an intelligent agent provided by an embodiment of the present invention;

[0057] Figure 5A schematic structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0059] In this application, the terms "comprises," "comprising," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not preclude the presence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0060] The embodiment of the present invention provides an observation method for an intelligent agent. The method can be applied to a business system in which an intelligent agent is deployed. The execution subject can be a processor of the system. The flow chart of the method is as follows: Figure 1 Shown, including:

[0061] S101: Determine sensor information corresponding to the intelligent agent to be observed; the sensor information includes sensor data corresponding to multiple information dimensions.

[0062] The methods provided by embodiments of the present invention can be applied to scenarios involving various intelligent agents making decisions, such as autonomous driving. During the operation of the intelligent agent, various sensor data used by the intelligent agent in making decisions can be collected in real time and used as sensor information corresponding to the intelligent agent. This sensor information includes multiple types of sensor data, i.e., sensor data from multiple information dimensions, which are determined by the decision-making logic of the intelligent agent. For example, in autonomous driving scenarios, data from various environmental sensors such as lidar, cameras, and infrared sensors can be obtained.

[0063] S102: Determine the internal state information corresponding to the agent.

[0064] In the method provided by the embodiment of the present invention, the current internal state of the intelligent agent can be obtained in real time, and the current internal state of the intelligent agent is used as the internal state information corresponding to the intelligent agent.

[0065] S103: Encode the sensor information and the internal state information respectively to obtain sensor encoding information corresponding to the sensor information and internal state encoding information corresponding to the internal state information.

[0066] In the method provided by the embodiments of the present invention, the sensor information of each dimension (i.e., sensor information) can be encoded using the encoder of the Transformer model, and the encoded data can be used as the sensor encoded information. The internal state of the agent (i.e., internal state information) can also be encoded using the encoder of the Transformer model, and the encoded data can be used as the internal state encoded information.

[0067] S104: Determine the decision result corresponding to the intelligent agent.

[0068] In the method provided in the embodiment of the present invention, the decision made by the intelligent agent based on various sensor data can be obtained, and the decision result corresponding to the intelligent agent can be obtained.

[0069] S105: Based on the sensor coding information, internal state coding information and decision results, construct a decision map corresponding to the intelligent agent.

[0070] In the method provided by an embodiment of the present invention, sensor coding information, internal state coding information and decision results are respectively used as node information in the form of a directed graph, and a graph recording the sequence and association between sensor data-internal state-decision results is constructed, and the graph is used as the decision graph corresponding to the intelligent agent.

[0071] S106: Extract features from the sensor information to obtain feature information corresponding to the sensor information.

[0072] In the method provided by the embodiments of the present invention, feature extraction models for extracting various types of features can be pre-established for each type of sensor data involved. For each type of sensor data in the sensor information, feature extraction can be performed using the corresponding feature extraction model, and the features extracted from each type of sensor data are used as the feature information corresponding to the sensor information.

[0073] S107: Based on the feature information, perform intention inference on the intelligent agent to obtain the intention inference result corresponding to the intelligent agent.

[0074] In the method provided by the embodiments of the present invention, an intention inference model can be pre-built. This model can infer the agent's intention based on the characteristics of each sensor data. During the observation process, the intention inference model can be applied to the characteristic information corresponding to the sensor data to infer the agent's intention. The results of the intention inference model are used as the inferred intention result for the agent.

[0075] S108: Determine the decision path information corresponding to the decision result based on the pre-built behavior evaluation map.

[0076] In the method provided by the embodiments of the present invention, a behavior evaluation map can be constructed in advance based on the behavior of the intelligent agent in different states, recording the decision basis and logic of the corresponding decision path. During the observation process, the decision path can be traced in the pre-constructed behavior evaluation map based on the corresponding decision result of the intelligent agent, and the decision path information corresponding to the decision result can be obtained, that is, the decision basis and logic in the decision path corresponding to the decision result.

[0077] S109: Based on the preset decision verification rules, the decision result is verified to obtain the decision verification result corresponding to the intelligent agent.

[0078] In the methods provided by embodiments of the present invention, decision verification rules can be pre-set based on actual business needs. These rules are used to verify whether decisions comply with corresponding rules. For example, in autonomous driving scenarios, the agent's decisions must comply with traffic regulations and safety principles. In service robot interaction scenarios, the agent's decisions must not violate ethical and legal norms. During the observation process, the pre-set decision verification rules can be applied to verify the agent's corresponding decision results. Specifically, the current decision results can be verified to determine whether they comply with the corresponding rules. The verification results are used as the agent's decision verification results. The decision verification results indicate whether the decision results comply with the corresponding rules.

[0079] S110: The decision map, intention inference results, decision path information and decision verification results are used as the observation results corresponding to the intelligent agent.

[0080] In the method provided by the embodiment of the present invention, the decision map, intention inference results, decision path information, and decision verification results corresponding to the intelligent agent are used as the current observation results. The observation results can be reflected to the user in real time or stored.

[0081] Based on the method provided by the embodiment of the present invention, when it is necessary to observe an intelligent agent, the sensor information corresponding to the intelligent agent to be observed can be determined; the sensor information includes sensor data corresponding to multiple information dimensions; the internal state information corresponding to the intelligent agent is determined; the sensor information and the internal state information are respectively encoded to obtain sensor coding information corresponding to the sensor information and internal state coding information corresponding to the internal state information; the decision result corresponding to the intelligent agent is determined; based on the sensor coding information, the internal state coding information and the decision result, a decision map corresponding to the intelligent agent is constructed; feature extraction is performed on the sensor information to obtain feature information corresponding to the sensor information; based on the feature information, the intention of the intelligent agent is inferred to obtain the intention inference result corresponding to the intelligent agent; based on a pre-constructed behavior evaluation map, the decision path information corresponding to the decision result is determined; based on a preset decision verification rule, the decision result is decision-verified to obtain the decision verification result corresponding to the intelligent agent; the decision map, the intention inference result, the decision path information and the decision verification result are used as the observation result corresponding to the intelligent agent. By applying the method provided in the embodiment of the present invention, a corresponding decision map can be constructed based on sensor data, the internal state information of the intelligent agent and the decision results, the intention of the intelligent agent can be inferred and the decision path can be traced back, and the decision results can be verified at the same time. Finally, an observation result including the decision map, intention inference results, decision path information and decision verification results is obtained. After the intelligent agent makes a decision, multi-dimensional observation of the intelligent agent can be performed in real time, which is conducive to timely discovery and handling of emergencies and improving the safety and reliability of the system.

[0082] exist Figure 1 Based on the method shown, in the method provided by the embodiment of the present invention, each information dimension includes an image information dimension and a voice information dimension. The process of extracting features from the sensor information mentioned in step S106 to obtain feature information corresponding to the sensor information includes:

[0083] For sensor data corresponding to the image information dimension, a pre-built image feature extraction model is applied to extract image features from the sensor data to obtain image features corresponding to the image information dimension; the image feature extraction model is a feature extraction model built based on a convolutional neural network;

[0084] In the method provided by the embodiment of the present invention, the sensor data of each information dimension in the sensor information includes sensor data corresponding to the image information dimension and sensor data corresponding to the voice information dimension, that is, includes image data and voice data.

[0085] In the method provided in an embodiment of the present invention, a feature extraction model for extracting image features is pre-built based on a convolutional neural network (CNN). This feature extraction model serves as the image feature extraction model. During the observation process, sensor data corresponding to the image information dimension in the sensor information, that is, image data, is input into the image feature extraction model. The image feature extraction model then extracts features from the corresponding image data, such as object features, and uses the extracted image features as the image features corresponding to the image information dimension.

[0086] For the sensor data corresponding to the voice information dimension, a pre-built voice feature extraction model is applied to extract voice features from the sensor data to obtain voice features corresponding to the voice information dimension; the voice feature extraction model is a feature extraction model built based on a long short-term memory network;

[0087] Image features and speech features are used as feature information corresponding to the intelligent agent.

[0088] In the method provided in an embodiment of the present invention, a feature extraction model for extracting speech features is pre-built based on a long short-term memory (LSTM) network. This feature extraction model serves as the speech feature extraction model. During the observation process, sensor data corresponding to the speech information block in the sensor information, that is, speech data, is input into the speech feature extraction model. The speech feature extraction model then extracts features from the corresponding speech data, such as temporal features of the speech. The extracted speech features are used as speech features corresponding to the speech information dimension. The speech features corresponding to the speech information dimension and the image features corresponding to the image information dimension are used as feature information corresponding to the intelligent agent.

[0089] exist Figure 1 On the basis of the method shown in the figure, in the method provided by the embodiment of the present invention, the process of performing intention inference on the agent based on the feature information mentioned in step S107 to obtain the intention inference result corresponding to the agent includes:

[0090] Apply a pre-built intention inference model to perform attention inference processing on the feature information to obtain the attention weights corresponding to the features of each region in the feature information; the intention inference model is a model built based on the attention mechanism.

[0091] In the method provided by the embodiments of the present invention, an intention inference model can be pre-built based on the attention mechanism. This intention inference model calculates the association weights between different regional features to identify the area where the agent's attention is focused. During the observation process, the agent's corresponding feature information can be input into the pre-built intention inference model, which then calculates the attention weights for each regional feature.

[0092] The attention weights corresponding to each regional feature are compared, and the regional feature with the relatively larger corresponding attention weight is used as the target regional feature.

[0093] In the method provided by the embodiment of the present invention, the attention weights corresponding to the various regional features calculated by the intention inference model can be compared, and the regional feature corresponding to the attention weight with the largest weight can be used as the target regional feature.

[0094] Based on the characteristics of the target area, the decision intention corresponding to the intelligent agent is determined, and the decision intention is used as the intention inference result corresponding to the intelligent agent.

[0095] In the method provided by the embodiment of the present invention, pre-stored prior knowledge or a pre-trained model can be applied to infer the target that the agent is concerned about in the target area features, thereby inferring the intention that the agent is concerned about, and using the inferred intention as the decision intention corresponding to the agent to obtain the intention inference result.

[0096] exist Figure 1 Based on the method shown, in the method provided in the embodiment of the present invention, the process of constructing the behavior evaluation map mentioned in step S108 includes:

[0097] Determine the various operating states corresponding to the agent.

[0098] Determine the value data set corresponding to each operating state; the value data set corresponding to each operating state includes the Q value corresponding to each action in the operating state.

[0099] Based on the value data sets corresponding to each operating state, a behavior evaluation map is constructed.

[0100] In the method provided by an embodiment of the present invention, an intelligent agent can be pre-programmed into different states within a reinforcement learning environment, and various state data of the intelligent agent during operation can be obtained. Based on these various state data, the Q-value corresponding to each action of the intelligent agent can be calculated. This is the estimated value of the long-term cumulative reward that the action can be expected to obtain in the current state. Based on the various data collected during the reinforcement learning process, the various possible operating states of the intelligent agent can be determined. For each operating state, the Q-value calculated for each action of the intelligent agent during the reinforcement learning process is obtained, and the Q-values ​​corresponding to each action are combined to form a value dataset corresponding to that operating state. A distribution map of Q-values ​​under different operating states can be constructed, with the motion state as the horizontal axis and the different actions and their corresponding Q-values ​​as the vertical axis. This distribution map can be used as a behavior evaluation map. The behavior evaluation map can intuitively display the value assessment of each action by the intelligent agent in different scenarios.

[0101] exist Figure 1 On the basis of the method shown in FIG. 1 , in the method provided in an embodiment of the present invention, the process of determining the decision path information corresponding to the decision result based on the pre-built behavior evaluation graph mentioned in step S108 includes:

[0102] Based on the behavior evaluation map, the behavior evaluation information corresponding to the internal state information is determined.

[0103] Based on the behavior evaluation information, the decision results are traced back to obtain the decision path information corresponding to the decision results.

[0104] In the method provided by embodiments of the present invention, during the observation process, based on the agent's current internal state, a path can be traced within the behavior evaluation graph to find behavior evaluation information that matches the current internal state, such as the Q-values ​​corresponding to various actions that match the current internal state. Within the behavior evaluation information, behavior evaluation data that matches the agent's current decision outcome is found. Based on the matching behavior evaluation data, the decision-making basis and logic for the current decision are determined, and the decision-making basis and logic for the current decision are used as the decision path information corresponding to the decision outcome. For example, in a game scenario, the agent controls a non-player character. When tracing the decision path, the Q-value distribution graph can be reviewed. For example, in a specific game scenario, the agent faces the decision of whether to attack or dodge an enemy. By examining the Q-values ​​of the "attack" and "dodge" actions in that state in the graph, if the "attack" action has a higher Q-value, it indicates that the agent believes that attacking in this state will bring a higher expected reward. This further clarifies the decision-making basis and logic for choosing the attack action, helping developers or analysts understand the agent's decision-making process and optimize the agent's behavior.

[0105] exist Figure 1On the basis of the method shown, the method provided in the embodiment of the present invention further includes:

[0106] Based on the decision verification results, the decision of the intelligent agent is optimized.

[0107] In the method provided in the embodiment of the present invention, the intelligent agent can be optimized based on whether the decision is compliant as fed back in the decision verification result. For example, if the decision verification result indicates that the decision is non-compliant, based on analysis, it is found that the non-compliant decision is caused by insufficient extraction of certain key environmental features. In this case, the feature extraction algorithm of the intelligent agent can be optimized to make the intelligent agent's decision more accurate.

[0108] exist Figure 1 On the basis of the method shown, the method provided in the embodiment of the present invention further includes:

[0109] During the observation of the intelligent agent, various data are transmitted based on the preset lightweight bus.

[0110] In the method provided by the embodiment of the present invention, a lightweight bus can be pre-set for data transmission. During the intelligent agent observation process, the lightweight bus is used to transmit and process data at each stage.

[0111] To better illustrate the method provided by the embodiment of the present invention, based on the agent observation methods provided in the previous embodiments, the embodiment of the present invention provides another agent observation method. The method provided by the embodiment of the present invention can be implemented using a dynamic perception matrix, a hierarchical state encoding module, and a lightweight observation bus.

[0112] The dynamic perception matrix is ​​primarily used to establish three-dimensional relationships between environmental sensors, the internal state of the AI ​​agent, and its decision outputs. It comprehensively perceives the information flow during the agent's interaction with the environment, providing a multi-dimensional understanding of the agent's behavior and decision-making. This three-dimensional relationship integrates information from diverse sources, breaking the limitations of traditional single-sensor data streams and eliminating observation blind spots. The dynamic perception matrix consists of three main components: the external perception layer, the state encoding layer, and the decision graph layer. The external perception layer collects data from various sensor types, such as lidar, cameras, and infrared sensors used in autonomous driving. The state encoding layer uses the Transformer model's encoder to encode the agent's internal state and sensor data, creating a corresponding feature representation and transforming complex states into a form that facilitates analysis. The decision graph layer constructs decision relationships in the form of a directed graph, recording the sequence and correlation of the agent's decisions. Specifically, based on the encoded information, the agent constructs a decision graph based on its decisions, such as acceleration and deceleration, to record the decision-making process and related information.

[0113] The hierarchical state coding module is mainly composed of five layers of observation models: physical sensing layer, feature extraction layer, intention inference layer, policy tracing layer, and ethical compliance layer. The functions of each layer are as follows:

[0114] Physical sensing layer: This layer directly acquires raw data from various sensors, such as point cloud data from lidar in autonomous driving, image data from cameras, and voice signals collected by microphones in service robots. This raw data provides the most direct perception of the agent's environment, but it is not interpretable and requires further processing.

[0115] Feature extraction layer: Deep learning models such as convolutional neural networks (CNNs) or long short-term memory networks (LSTMs) are used to extract features from the raw data from the physical sensing layer. For example, for image data, CNNs are used to extract object features; for speech data, LSTMs are used to extract speech temporal features. These features can more effectively represent the intrinsic information of the data and provide a foundation for subsequent analysis.

[0116] The intention inference layer uses an attention mechanism to process the features obtained by the feature extraction layer. Attention weights are visualized to infer the agent's intention. Specifically, by calculating the correlation weights between image features in different regions, the agent can highlight features that are more critical to the current decision. For example, in autonomous driving scenarios, by analyzing the agent's attention to features in different regions, the agent's attention targets, such as pedestrians or vehicles ahead, can be inferred. For example, when processing images of the road ahead, the attention mechanism will cause the agent to focus more on features in areas where pedestrians or vehicles are likely to appear. These calculated attention weights are visualized to intuitively illustrate the agent's attention to different image features. Based on this visualization, if a region has a high attention weight, combined with prior knowledge or a trained model, it can be inferred that the agent's attention target is located in that region, such as whether it is a pedestrian or vehicle ahead. This helps clarify the agent's attention intent and provide a basis for subsequent decision-making.

[0117] Decision Tracing Layer: During decision-making processes like reinforcement learning, a Q-value distribution map (i.e., the behavior evaluation map in the previous example) is constructed to record the estimated value of different actions taken by the agent in different states. By analyzing the distribution and changes in Q-values, the agent's decision path can be traced, clarifying the basis and logic behind the decision. For example, in the decision-making process of non-player characters in games, it is possible to clearly see how they choose actions based on the environment state and reward mechanism.

[0118] Ethics and Compliance Layer: This layer pre-sets a series of ethical and compliance rules to verify the decisions of intelligent agents. For example, in autonomous driving, this layer ensures that the agent's decisions comply with traffic regulations and safety principles. In service robot-user interactions, this layer ensures that the agent's responses do not violate ethical and legal norms.

[0119] In the method provided by an embodiment of the present invention, the data generated by the dynamic perception matrix and the hierarchical state coding module at each stage are transmitted using a lightweight observation bus. Deployed using an edge computing architecture, the lightweight observation bus is a dedicated data channel. It places data processing and analysis as close to the data source as possible, reducing data transmission delays. By optimizing data transmission paths and employing efficient data processing algorithms, rapid transmission and processing of observation data is achieved. The design and deployment scheme for a lightweight observation bus based on an edge computing architecture, including technical details for optimizing data transmission paths, efficient data processing algorithms, and reducing observation data transmission delays, ensures real-time observation.

[0120] Specifically, the example diagram of the agent observation process can be as follows Figure 2 As shown in Figure 1, the hierarchical state encoding module consists of five layers: L1-L5, namely the physical sensing layer (L1), feature extraction layer (L2), intent inference layer (L3), decision tracing layer (L4), and ethical compliance layer (L5). These five layers form a complete encoding system, providing progressively deeper processing and analysis of data, transforming raw data into interpretable semantic representations. Figure 2 This paper describes the interaction between data processed by the hierarchical state encoding module, the dynamic perception matrix, and the agent's decision engine. The decision engine is equipped with a corresponding policy rule base, a real-time inference engine, and a feedback learner. The dynamic perception matrix collects environmental sensor data, encodes the agent's internal state, and constructs a decision graph. The generated data serves as input to the hierarchical state encoding module. After five layers of encoding, the final output provides key information to the decision engine, assisting it in making the agent's decision. This completes the entire process from environmental perception to state encoding to decision making.

[0121] The state encoding process of the hierarchical state encoding module can be as follows Figure 3As shown, the physical sensing layer collects raw sensor data. The feature extraction layer extracts features from this data using a CNN or LSTM model. The intent inference layer infers the agent's intent using attention weights, and the decision tracing layer performs decision tracing. Finally, the ethics and compliance layer verifies the agent's decision against pre-set rules to determine whether it complies with pre-set ethical, legal, and other regulations. If any non-compliance is detected, the ethics and compliance layer generates corresponding feedback. This feedback is then passed back to the feature extraction layer via a hierarchical state encoding module. Once back to the feature extraction layer, the layer can use this compliance feedback to perform more targeted feature extraction on subsequent raw data input (from the physical sensing layer). For example, if a decision in a particular decision scenario is found to be non-compliant due to insufficient extraction of certain key environmental features, the feature extraction layer can adjust the feature extraction algorithm or parameters to prioritize these key features. On the other hand, compliance feedback can also prompt the feature extraction layer to re-evaluate and screen the extracted features, eliminate interfering features that may lead to non-compliant decisions, retain or strengthen effective features related to compliance decisions, thereby optimizing the entire intelligent agent's process from data processing to decision generation, making subsequent decisions more in line with ethical compliance requirements.

[0122] The method provided in the embodiment of the present invention can adopt a compatibility design that is compatible with existing AI development frameworks. Specifically, for interface design, the interface specifications of existing AI development frameworks (such as ROS and TensorFlow) can be analyzed, and general communication protocols and interaction modes can be extracted to design a unified, concise and extensible interface to achieve efficient connection between this technology and each framework and reduce the difficulty of access. For data format conversion, the data format standards supported by existing AI development frameworks can be studied, and data parsing and reconstruction algorithms can be constructed based on the characteristics of the data generated by this method to achieve accurate conversion of data formats and ensure the compatibility and availability of data between different frameworks. For functional adaptation, the functional system and application scenarios of existing AI development frameworks are deeply analyzed. Based on the functional characteristics of this technology, a modular design is adopted to develop adaptation plug-ins or components so that the functions implemented by this method are organically integrated with the functions of existing frameworks to meet different development needs.

[0123] Based on the method provided in the embodiment of the present invention, by constructing a dynamic perception matrix, adopting a hierarchical state coding module, and deploying a lightweight observation bus, observability of the entire life cycle of the AI ​​agent can be achieved, thereby improving the security, reliability, and performance of the AI ​​agent in different application scenarios.

[0124] and Figure 1 Corresponding to the observation method of an intelligent agent shown in FIG, an embodiment of the present invention further provides an observation device for an intelligent agent, Figure 1 The specific implementation of the method shown in is shown in the structural diagram. Figure 4 Shown, including:

[0125] The first determining unit 201 is used to determine the sensor information corresponding to the intelligent agent to be observed; the sensor information includes sensor data corresponding to multiple information dimensions.

[0126] The second determining unit 202 is used to determine the internal state information corresponding to the agent.

[0127] The information encoding unit 203 is configured to encode the sensor information and the internal state information respectively to obtain sensor encoding information corresponding to the sensor information and internal state encoding information corresponding to the internal state information.

[0128] The third determining unit 204 is used to determine the decision result corresponding to the agent.

[0129] The graph construction unit 205 is used to construct a decision graph corresponding to the intelligent agent based on the sensor coding information, the internal state coding information and the decision result.

[0130] The feature extraction unit 206 is used to extract features from the sensor information to obtain feature information corresponding to the sensor information.

[0131] The intention inference unit 207 is used to infer the intention of the agent based on the feature information and obtain the intention inference result corresponding to the agent.

[0132] The fourth determining unit 208 is configured to determine decision path information corresponding to the decision result based on a pre-built behavior evaluation graph.

[0133] The decision verification unit 209 is used to perform decision verification on the decision result based on a preset decision verification rule to obtain a decision verification result corresponding to the intelligent agent.

[0134] The fifth determination unit 210 is used to use the decision map, intention inference results, decision path information and decision verification results as observation results corresponding to the intelligent agent.

[0135] By using the device provided by the embodiment of the present invention, a corresponding decision map can be constructed based on sensor data, the internal state information of the intelligent agent and the decision results, the intention of the intelligent agent can be inferred and the decision path can be traced back, and the decision results can be verified at the same time. Finally, an observation result including the decision map, intention inference results, decision path information and decision verification results is obtained. After the intelligent agent makes a decision, multi-dimensional observation of the intelligent agent can be performed in real time, which is conducive to timely discovery and handling of emergencies and improving the safety and reliability of the system.

[0136] exist Figure 4Based on the device shown, the device provided by the embodiment of the present invention can be further expanded into multiple units. The functions of each unit can be found in the description of the various embodiments provided in the previous text for the observation method of the intelligent agent, and no further examples will be given here.

[0137] An embodiment of the present invention further provides a storage medium, which includes stored instructions, wherein when the instructions are executed, the device where the storage medium is located is controlled to execute the above-mentioned intelligent agent observation method.

[0138] The embodiment of the present invention further provides an electronic device, the structural diagram of which is shown in FIG. Figure 5 As shown, the system specifically includes a memory 301 and one or more instructions 302, wherein the one or more instructions 302 are stored in the memory 301 and are configured to be executed by one or more processors 303 to perform the following operations:

[0139] Determine the sensor information corresponding to the intelligent agent to be observed; the sensor information includes sensor data corresponding to multiple information dimensions.

[0140] Determine the internal state information corresponding to the agent.

[0141] The sensor information and the internal state information are encoded respectively to obtain sensor encoding information corresponding to the sensor information and internal state encoding information corresponding to the internal state information.

[0142] Determine the decision result corresponding to the agent.

[0143] Based on the sensor encoding information, internal state encoding information and decision results, a decision graph corresponding to the intelligent agent is constructed.

[0144] Perform feature extraction on the sensor information to obtain feature information corresponding to the sensor information.

[0145] Based on the feature information, the agent's intention is inferred to obtain the agent's corresponding intention inference result.

[0146] Determine the decision path information corresponding to the decision result based on the pre-built behavior evaluation map;

[0147] Based on the preset decision verification rules, the decision results are verified to obtain the decision verification results corresponding to the intelligent agent.

[0148] The decision map, intention inference results, decision path information and decision verification results are used as the observation results corresponding to the intelligent agent.

[0149] Each embodiment in this specification is described in a progressive manner. The same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple. For relevant parts, refer to the partial description of the method embodiment. The system and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0150] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0151] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An observation method for an intelligent agent, characterized in that: include: Determine the sensor information corresponding to the intelligent agent to be observed; The sensor information includes sensor data corresponding to multiple information dimensions; Determining internal state information corresponding to the agent; Encode the sensor information and the internal state information respectively to obtain sensor encoding information corresponding to the sensor information and internal state encoding information corresponding to the internal state information; Determining a decision result corresponding to the agent; Constructing a decision graph corresponding to the agent based on the sensor encoding information, the internal state encoding information and the decision result; Performing feature extraction on the sensor information to obtain feature information corresponding to the sensor information; Based on the feature information, performing intention inference on the agent to obtain an intention inference result corresponding to the agent; Determine the decision path information corresponding to the decision result based on a pre-built behavior evaluation map; Based on a preset decision verification rule, the decision result is verified to obtain a decision verification result corresponding to the intelligent agent; The decision map, the intention inference result, the decision path information and the decision verification result are used as the observation results corresponding to the intelligent agent.

2. The method for observing an intelligent agent according to claim 1, wherein: Each of the information dimensions includes an image information dimension and a voice information dimension. The feature extraction of the sensor information to obtain feature information corresponding to the sensor information includes: For the sensor data corresponding to the image information dimension, applying a pre-built image feature extraction model to perform image feature extraction on the sensor data to obtain image features corresponding to the image information dimension; the image feature extraction model is a feature extraction model built based on a convolutional neural network; For the sensor data corresponding to the voice information dimension, a pre-built voice feature extraction model is applied to the sensor data to extract voice features, thereby obtaining voice features corresponding to the voice information dimension; the voice feature extraction model is a feature extraction model built based on a long short-term memory network; The image features and the voice features are used as feature information corresponding to the agent.

3. The method for observing an intelligent agent according to claim 1, wherein: The performing intention inference on the agent based on the feature information to obtain an intention inference result corresponding to the agent includes: Applying a pre-built intention inference model to perform attention inference processing on the feature information to obtain attention weights corresponding to the features of each region in the feature information; the intention inference model is a model built based on the attention mechanism; Compare the attention weights corresponding to the respective regional features, and use the regional feature with the relatively larger corresponding attention weight as the target regional feature; Based on the target area characteristics, the decision intention corresponding to the intelligent agent is determined, and the decision intention is used as the intention inference result corresponding to the intelligent agent.

4. The method for observing an intelligent agent according to claim 1, wherein: The process of constructing the behavior assessment map includes: Determining each operating state corresponding to the agent; Determine a value data set corresponding to each of the operating states; the value data set corresponding to each of the operating states includes a Q value corresponding to each action in the operating state; The behavior evaluation map is constructed based on the value data sets corresponding to each of the operating states.

5. The method for observing an intelligent agent according to claim 1, wherein: The determining of decision path information corresponding to the decision result based on the pre-built behavior evaluation graph includes: Determining behavior evaluation information corresponding to the internal state information based on the behavior evaluation map; Based on the behavior evaluation information, the decision result is traced back to obtain decision path information corresponding to the decision result.

6. The method for observing an intelligent agent according to claim 1, wherein: Also includes: Based on the decision verification result, the decision of the intelligent agent is optimized.

7. The method for observing an intelligent agent according to claim 1, wherein: Also includes: During the process of observing the intelligent agent, various data are transmitted based on a preset lightweight bus.

8. An observation device for an intelligent agent, characterized in that: include: A first determining unit, configured to determine sensor information corresponding to the intelligent agent to be observed; The sensor information includes sensor data corresponding to multiple information dimensions; A second determining unit, configured to determine internal state information corresponding to the agent; an information encoding unit, configured to encode the sensor information and the internal state information respectively, to obtain sensor encoding information corresponding to the sensor information and internal state encoding information corresponding to the internal state information; A third determining unit, configured to determine a decision result corresponding to the agent; A graph construction unit, configured to construct a decision graph corresponding to the agent based on the sensor encoding information, the internal state encoding information, and the decision result; a feature extraction unit, configured to extract features from the sensor information to obtain feature information corresponding to the sensor information; an intention inference unit, configured to infer the intention of the agent based on the feature information and obtain an intention inference result corresponding to the agent; a fourth determining unit, configured to determine decision path information corresponding to the decision result based on a pre-built behavior evaluation graph; A decision verification unit, configured to perform decision verification on the decision result based on a preset decision verification rule to obtain a decision verification result corresponding to the intelligent agent; The fifth determination unit is used to use the decision map, the intention inference result, the decision path information and the decision verification result as the observation results corresponding to the intelligent agent.

9. A storage medium, characterized in that: The storage medium includes stored instructions, wherein when the instructions are executed, the device where the storage medium is located is controlled to execute the observation method of the intelligent agent according to any one of claims 1 to 7.

10. An electronic device, characterized in that: The device comprises a memory and one or more instructions, wherein the one or more instructions are stored in the memory and configured to be executed by one or more processors to perform the observation method of the intelligent agent as described in any one of claims 1 to 7.

Images