Dynamic knowledge graph-based domain agent self-adaptive decision method and device, equipment and storage medium

Abstract

This application discloses a method, apparatus, device, and storage medium for adaptive decision-making by a domain intelligent agent based on a dynamic knowledge graph. The method includes: real-time acquisition of raw multi-source data from multiple heterogeneous data sources, and performing three-dimensional indexing preprocessing on the raw multi-source data at the edge to obtain preprocessed multi-source data; using incremental knowledge extraction technology based on the preprocessed multi-source data to dynamically identify and extract entity information, relationships, and event information, and constructing a dynamic knowledge graph; obtaining the current decision task, and searching for candidate decision paths on the dynamic knowledge graph based on the current decision task using reinforcement learning and path reasoning algorithms; using a policy gradient algorithm combined with a domain reward function to optimize and rank the candidate decision paths, generating the optimal decision strategy; and feeding back the optimal decision strategy to the task initiator to guide the domain intelligent agent in completing adaptive decision-making. This application can effectively improve the real-time performance and accuracy of the domain intelligent agent's decision-making.

Description

Technical Field

[0001] This application relates to the field of knowledge graph technology, and in particular to adaptive decision-making methods, devices, equipment and storage media for domain intelligent agents based on dynamic knowledge graphs. Background Technology

[0002] With the development of big data and artificial intelligence technologies, domain-specific intelligent agents are being applied more and more widely in fields such as transportation, industry, and healthcare. Traditional domain-specific intelligent agent decision-making methods typically rely on static knowledge bases and pre-set rules, making it difficult to adapt to complex and ever-changing real-world environments and failing to fully utilize multi-source heterogeneous data, resulting in insufficient accuracy and real-time performance in decision-making. Knowledge graphs, as an effective knowledge representation and management tool, can fuse multi-source heterogeneous data into semantic networks, providing intelligent agents with rich background knowledge. However, existing knowledge graph construction methods are mostly static or semi-static, making it difficult to reflect the dynamic changes in domain knowledge in real time, thus limiting the adaptive decision-making capabilities of intelligent agents in complex environments. Summary of the Invention

[0003] The main objective of this application is to provide a method, apparatus, device, and storage medium for adaptive decision-making of domain intelligent agents based on dynamic knowledge graphs, aiming to solve the technical problems of low decision-making accuracy, poor real-time performance, and weak adaptability of domain intelligent agents.

[0004] To achieve the above objectives, this application proposes a domain agent adaptive decision-making method based on dynamic knowledge graphs. The domain agent adaptive decision-making method based on dynamic knowledge graphs includes: Raw multi-source data from multiple heterogeneous data sources is collected in real time, and the raw multi-source data is preprocessed with a three-dimensional index of time, space and semantics at the edge to obtain preprocessed multi-source data. Based on the preprocessed multi-source data, incremental knowledge extraction technology is used to dynamically identify and extract entity information, relationships and event information, and a dynamic knowledge graph is constructed based on the entity information, relationships and event information. Obtain the current decision task, and search for candidate decision paths on the dynamic knowledge graph based on the current decision task using reinforcement learning and path reasoning algorithms; The candidate decision paths are optimized and ranked using a policy gradient algorithm combined with a domain reward function to generate the optimal decision strategy. The optimal decision-making strategy is fed back to the task initiator to guide the domain agent in making adaptive decisions.

[0005] In one embodiment, the real-time acquisition of raw multi-source data from multiple heterogeneous data sources, and the edge-side preprocessing of the raw multi-source data with a three-dimensional temporal-spatial-semantic index to obtain preprocessed multi-source data, includes: A distributed data acquisition framework is used to perform real-time synchronous acquisition of multi-source heterogeneous data to obtain raw multi-source data. A spatiotemporal semantic coding model is deployed through edge computing nodes, and the original multi-source data is indexed and encoded in terms of time, space and semantic dimensions through the spatiotemporal semantic coding model to construct a three-dimensional index structure that includes temporal features, spatial distribution and semantic association. Based on the three-dimensional index structure, the original multi-source data is retrieved and located to extract multi-source data fragments related to the current decision-making task. A lightweight data cleaning rule engine is used to perform outlier filtering, missing value filling and unit unification preprocessing on the multi-source data fragments at the edge, generating preprocessed multi-source data.

[0006] In one embodiment, the incremental knowledge extraction technique is used based on the preprocessed multi-source data to dynamically identify and extract entity information, relationships, and event information, and a dynamic knowledge graph is constructed based on the entity information, relationships, and event information, including: The preprocessed multi-source data stream is sliced ​​in three dimensions according to time window, spatial grid, and semantic topic to form independently evolving slice data packets; Emergence is calculated for each slice data packet to obtain the emergence value of each slice data packet, wherein the emergence value is used to measure the probability and importance of the occurrence of new entities, new relationships or new events in the slice data packet; Slice data packets whose emergence values ​​reach a preset emergence threshold are used as entity incubation candidate sets; Based on the characteristic distribution of data packets in the entity incubation candidate set, an adaptive clustering algorithm is used to identify potential entity clusters, and feature extraction and pattern matching are performed on each potential entity cluster to obtain entity information; The entity information is subjected to relationship pulse sniffing to identify the relationships between entities; The entity information and the relationship are used as nodes to form a transient cloud, and the high-density sub-clouds in the transient cloud are identified by the cloud cluster cohesion index. The high-density sub-cloud is mapped to potential events, and the confidence of the potential events is evaluated through a confidence ripple feedback mechanism to determine the event information; A topology-temporal dual-layer signature algorithm is used to uniquely identify the entity information, the relationship, and the event information to construct a dynamic knowledge graph.

[0007] In one embodiment, obtaining the current decision task and searching for candidate decision paths on the dynamic knowledge graph based on the current decision task using reinforcement learning and path reasoning algorithms includes: The dynamic knowledge graph is subjected to multi-scale compression and subgraph partitioning to generate cascaded subgraphs adapted to different decision-making tasks; Obtain the current decision task, perform a formal transformation on the current decision task, and generate a graph query; Reinforcement learning and path reasoning algorithms are used to search for candidate decision paths on the cascaded subgraphs based on the graph query.

[0008] In one embodiment, the step of performing multi-scale compression and subgraph partitioning on the dynamic knowledge graph to generate cascaded subgraphs adapted to different decision-making tasks includes: The dynamic knowledge graph is initially compressed using a graph clustering algorithm to generate multiple subgraph modules; By setting different compression granularity parameters, the subgraph module is recursively compressed twice to form a multi-layered cascaded subgraph structure; Perform a three-dimensional density pulse scan on the nodes in the dynamic knowledge graph to generate a density pulse map; Based on the density peak and density valley values ​​in the density pulse map, the multi-scale segmentation boundary is determined. The multi-level cascaded subgraph structure is segmented and adjusted according to the multi-scale segmentation boundary to generate cascaded subgraphs adapted to different decision-making tasks.

[0009] In one embodiment, the step of using reinforcement learning and path reasoning algorithms to search for candidate decision paths on the cascaded subgraph based on the graph query includes: The graph query is transformed into a state representation in a reinforcement learning environment, wherein the state representation includes information on the current query node, visited paths, and remaining query requirements. In a reinforcement learning environment, an action space is defined, which includes all feasible edges originating from the current query node and their corresponding target nodes. Based on state representation and action space, an action selection policy is generated using a policy network, wherein the action selection policy is used to select the optimal action at each decision step to maximize long-term reward. The state representation is updated according to the action selection strategy to obtain the updated state representation; Path exploration is performed on the cascaded subgraph based on the updated state representation and action selection strategy. After each action jump is completed, the remaining query requirement information is updated. When the remaining query requirements meet the preset termination conditions or the maximum number of exploration steps is reached, path exploration stops, and the set of paths already explored is used as candidate decision paths.

[0010] In one embodiment, the step of employing a policy gradient algorithm, combined with a domain reward function, to optimize and rank the candidate decision paths and generate an optimal decision policy includes: Define a domain reward function, which is used to assign a corresponding reward value to each candidate decision path based on the effectiveness, timeliness, resource consumption, and domain-specific constraints of the decision path; The total reward value for each candidate decision path is calculated based on the domain reward function, wherein the total reward value is the sum or weighted sum of the reward values ​​of each node on the candidate decision path; The policy gradient algorithm is used to optimize the candidate decision path by gradient ascent based on the total reward value, and the action selection probability distribution of each node on the path is adjusted to obtain the optimized action selection probability distribution. Based on the optimized action selection probability distribution, the candidate decision paths are reordered, and the candidate decision path with the highest total reward value is selected as the optimal decision strategy.

[0011] Furthermore, to achieve the above objectives, this application also proposes a domain intelligent agent adaptive decision-making device based on dynamic knowledge graphs. The domain intelligent agent adaptive decision-making device based on dynamic knowledge graphs includes: The preprocessing module is used to collect raw multi-source data from multiple heterogeneous data sources in real time, and perform three-dimensional index preprocessing of the raw multi-source data in terms of time, space and semantics on the edge side to obtain preprocessed multi-source data. The construction module is used to dynamically identify and extract entity information, relationship and event information based on the preprocessed multi-source data using incremental knowledge extraction technology, and to construct a dynamic knowledge graph based on the entity information, relationship and event information; The search module is used to obtain the current decision task and, based on the current decision task, search for candidate decision paths on the dynamic knowledge graph using reinforcement learning and path reasoning algorithms. The ranking module is used to optimize and rank the candidate decision paths using a policy gradient algorithm combined with a domain reward function, thereby generating the optimal decision strategy. The feedback module is used to feed back the optimal decision-making strategy to the task initiator to guide the domain agent to complete adaptive decision-making.

[0012] Furthermore, to achieve the above objectives, this application also proposes a domain intelligent agent adaptive decision-making device based on dynamic knowledge graphs. The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. The computer program is configured to implement the steps of the domain intelligent agent adaptive decision-making method based on dynamic knowledge graphs as described above.

[0013] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the domain intelligent agent adaptive decision-making method based on dynamic knowledge graph as described above.

[0014] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the domain intelligent agent adaptive decision-making method based on dynamic knowledge graph as described above.

[0015] This application proposes one or more technical solutions that involve real-time acquisition of raw multi-source data from multiple heterogeneous data sources, and edge-side preprocessing of the raw multi-source data using a three-dimensional temporal-spatial-semantic index to obtain preprocessed multi-source data. Based on the preprocessed multi-source data, incremental knowledge extraction technology is employed to dynamically identify and extract entity information, relationships, and event information, and a dynamic knowledge graph is constructed based on the entity information, relationships, and event information. The current decision task is obtained, and based on the current decision task, reinforcement learning and path reasoning algorithms are used to search for candidate decision paths on the dynamic knowledge graph. A policy gradient algorithm, combined with a domain reward function, is used to optimize and rank the candidate decision paths to generate an optimal decision strategy. The optimal decision strategy is fed back to the task initiator to guide the domain agent in completing adaptive decision-making. Through the above methods, by real-time acquisition and preprocessing of multi-source heterogeneous data, constructing a dynamic knowledge graph, and performing adaptive learning and decision-making based on the graph, the real-time performance and accuracy of the domain agent's decision-making can be effectively improved. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating an embodiment of the adaptive decision-making method for domain intelligent agents based on dynamic knowledge graphs in this application. Figure 2 This is a flowchart illustrating Embodiment 2 of the domain intelligent agent adaptive decision-making method based on dynamic knowledge graphs in this application. Figure 3This is a schematic diagram of the module structure of the domain intelligent agent adaptive decision-making device based on dynamic knowledge graph according to an embodiment of this application; Figure 4 This is a schematic diagram of the hardware operating environment of the adaptive decision-making device for a domain intelligent agent based on a dynamic knowledge graph, as described in this application embodiment.

[0019] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0021] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0022] The main solution of this application embodiment is as follows: Real-time acquisition of raw multi-source data from multiple heterogeneous data sources, and edge-side preprocessing of the raw multi-source data using a three-dimensional index of time, space, and semantics to obtain preprocessed multi-source data; Based on the preprocessed multi-source data, incremental knowledge extraction technology is used to dynamically identify and extract entity information, relationships, and event information, and a dynamic knowledge graph is constructed based on the entity information, relationships, and event information; The current decision task is obtained, and based on the current decision task, reinforcement learning and path reasoning algorithms are used to search for candidate decision paths on the dynamic knowledge graph; A policy gradient algorithm, combined with a domain reward function, is used to optimize and rank the candidate decision paths to generate the optimal decision strategy; The optimal decision strategy is fed back to the task initiator to guide the domain agent in completing adaptive decision-making.

[0023] This application provides a solution that constructs a dynamic knowledge graph by real-time acquisition and preprocessing of multi-source heterogeneous data, and performs adaptive learning and decision-making based on the graph, which can effectively improve the real-time performance and accuracy of domain intelligent agents' decision-making.

[0024] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of performing the above functions, such as a domain intelligent agent adaptive decision-making device based on dynamic knowledge graphs. The following description uses a domain intelligent agent adaptive decision-making device based on dynamic knowledge graphs as an example to illustrate this embodiment and the subsequent embodiments.

[0025] Based on this, embodiments of this application provide a domain-specific adaptive decision-making method for intelligent agents based on dynamic knowledge graphs, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the adaptive decision-making method for domain intelligent agents based on dynamic knowledge graphs in this application.

[0026] In this embodiment, the domain-based intelligent agent adaptive decision-making method based on dynamic knowledge graphs includes steps S10~S50: Step S10: Collect raw multi-source data from multiple heterogeneous data sources in real time, and perform time-space-semantic three-dimensional index preprocessing on the raw multi-source data at the edge to obtain preprocessed multi-source data.

[0027] It should be noted that data acquisition devices deployed at the edge can capture raw, multi-source data from different data sources (such as sensors, databases, log files, etc.) in real time. This data undergoes preliminary temporal-spatial-semantic three-dimensional indexing preprocessing at the edge, aiming to improve the efficiency and accuracy of data processing and lay the foundation for subsequent knowledge extraction and graph construction.

[0028] It is understood that this application can be applied to multiple fields, such as transportation, healthcare, finance, and industrial manufacturing. These fields typically involve a large number of heterogeneous data sources and require real-time and accurate processing and analysis of these data to support decision-making. This implementation will be illustrated using the transportation field as an example.

[0029] In the transportation sector, the solution presented in this application can be applied to intelligent traffic management systems. It involves real-time acquisition of raw, multi-source data from multiple heterogeneous data sources, such as traffic cameras, sensors, and GPS devices, and performing three-dimensional temporal-semantic indexing preprocessing on this raw, multi-source data at the edge. For example, video data captured by traffic cameras can extract information such as vehicle position, speed, and direction of travel, and index it using timestamps and geographic location information; traffic flow data collected by sensors can be aggregated and indexed according to time periods and geographic locations; and vehicle trajectory data provided by GPS devices can be segmented and semantically labeled for subsequent analysis and decision-making.

[0030] In one feasible implementation, step S10 may include: using a distributed data acquisition framework to perform real-time synchronous acquisition of multi-source heterogeneous data to obtain raw multi-source data; deploying a spatiotemporal semantic coding model through edge computing nodes, and using the spatiotemporal semantic coding model to index and encode the raw multi-source data in terms of time dimension, spatial dimension, and semantic dimension, constructing a three-dimensional index structure containing temporal features, spatial distribution, and semantic association; retrieving and locating the raw multi-source data based on the three-dimensional index structure, and extracting multi-source data fragments related to the current decision-making task; and using a lightweight data cleaning rule engine to perform outlier filtering, missing value filling, and unit unification preprocessing on the multi-source data fragments at the edge side to generate preprocessed multi-source data.

[0031] It should be noted that the distributed data acquisition framework has high scalability and fault tolerance, and can cope with the real-time acquisition needs of large-scale heterogeneous data sources, ensuring the integrity and timeliness of data. Real-time synchronous acquisition of multi-source heterogeneous data through the distributed data acquisition framework can greatly improve the efficiency and reliability of data acquisition.

[0032] Deploying spatiotemporal semantic coding models via edge computing nodes refers to placing these models on edge computing nodes. Leveraging their proximity to the data source reduces data transmission latency and improves data processing efficiency. Edge computing nodes can receive raw, multi-source data from a distributed data acquisition framework in real time and immediately index and encode this data using the spatiotemporal semantic coding model, considering time, space, and semantic dimensions. This deployment method makes data processing faster and more efficient, meeting the needs of real-time decision-making. Furthermore, since edge computing nodes typically possess certain computing and storage capabilities, they can complete some data processing tasks locally, reducing the burden on the central server.

[0033] The purpose of building a three-dimensional index structure is to manage and query multi-source data more efficiently. This three-dimensional index structure not only considers the temporal order and spatial location of the data, but also deeply explores the semantic relationships between the data, making subsequent data retrieval and positioning more accurate and faster. For example, in the field of transportation, this three-dimensional index structure can quickly locate traffic events that occur in a specific time period and at a specific location, as well as related vehicle, pedestrian and other entity information, providing strong support for subsequent decision-making.

[0034] The three-dimensional index structure is constructed through index encoding in three dimensions: time, space, and semantics. The time dimension index encoding sorts and segments the raw multi-source data based on timestamps to track data trends over time. For example, in traffic flow monitoring, the time dimension index can be accurate to the minute or second, capturing real-time fluctuations in traffic flow. The spatial dimension index encoding utilizes Geographic Information System (GIS) technology to associate data with specific geographical locations. By assigning latitude and longitude coordinates or grid codes to each data point, spatial distribution maps can be constructed, visually displaying the data's performance in different regions. This is particularly important in traffic management, as traffic conditions can vary drastically across different areas, and spatial indexing helps quickly locate problem areas. The semantic dimension index encoding focuses on uncovering the meaning and relationships behind the data. Through natural language processing (NLP) techniques and domain knowledge graphs, semantic analysis can be performed on text data to extract key entities, events, and relationships. For example, when processing video data captured by traffic cameras, semantic indexing can identify vehicle types, driving states (such as speeding and illegal lane changes), and interactions with the surrounding environment, supporting the automatic identification and early warning of traffic violations.

[0035] Based on the constructed three-dimensional index structure, it is possible to efficiently retrieve and locate multi-source data fragments closely related to the current decision-making task. These data fragments may include traffic flow data within a specific time period, vehicle trajectory information at a specific location, or all entity and relationship data related to a traffic event. Through precise data retrieval, the key information needed for decision-making can be quickly focused on, improving decision-making efficiency and accuracy.

[0036] At the edge, a lightweight data cleaning rule engine is used to preprocess the retrieved multi-source data fragments. This step aims to remove noise and outliers, fill in missing values, and standardize data units to ensure data quality and consistency. The lightweight design means the engine can run efficiently on resource-constrained edge devices without placing an excessive computational burden on them. Through data cleaning and preprocessing, a cleaner and more accurate dataset is obtained, laying a solid foundation for subsequent knowledge extraction and graph construction.

[0037] Step S20: Based on the preprocessed multi-source data, incremental knowledge extraction technology is used to dynamically identify and extract entity information, relationship and event information, and a dynamic knowledge graph is constructed based on the entity information, relationship and event information.

[0038] It's important to note that incremental knowledge extraction is a technique that continuously updates and expands the knowledge graph as new data is added. Unlike traditional batch knowledge extraction, incremental knowledge extraction does not require reprocessing all data; instead, it extracts and updates only the newly added or changed data, thus significantly improving the efficiency and real-time performance of knowledge extraction.

[0039] In this application, incremental knowledge extraction technology is applied to dynamically identify and extract entity information, relationships, and event information from preprocessed multi-source data. This information forms the basis for constructing a dynamic knowledge graph. Through incremental knowledge extraction, new entities, relationships, and events in the data can be captured in real time and integrated into the existing knowledge graph, enabling the knowledge graph to continuously update and evolve with time and data changes. For example, in the transportation field, incremental knowledge extraction technology can identify newly emerging traffic participants (such as vehicles and pedestrians), new traffic events (such as traffic accidents and road construction), and their relationships (such as the relationship between vehicles and accidents, or between vehicles and construction areas) in real time. After this information is extracted, it is immediately added to the dynamic knowledge graph, providing the latest data support for subsequent decision-making. In addition, incremental knowledge extraction technology can also handle dynamic changes in data. For example, when the state of a traffic participant changes (such as vehicle speed or direction), incremental knowledge extraction technology can capture these changes in a timely manner and update the relevant information in the knowledge graph. This ability to dynamically update knowledge graphs enables them to more accurately reflect the actual situation in the real world, providing more reliable data for decision-making.

[0040] In one feasible implementation, step S20 may include: slicing the preprocessed multi-source data stream into three dimensions according to time window-spatial grid-semantic topic to form independently evolving slice data packets; calculating the emergence degree of each slice data packet to obtain an emergence degree value, wherein the emergence degree value is used to measure the probability and importance of new entities, new relationships, or new events appearing in the slice data packet; selecting slice data packets whose emergence degree values ​​reach a preset emergence degree threshold as entity incubation candidate sets; and identifying potential entities using an adaptive clustering algorithm based on the feature distribution of data packets in the entity incubation candidate set. Clusters are identified, and feature extraction and pattern matching are performed on each potential entity cluster to obtain entity information. Relationship pulse sniffing is performed on the entity information to identify the relationships between entities. A transient graph cloud is constructed using the entity information and the relationships as nodes, and high-density sub-clouds in the transient graph cloud are identified through the cluster cloud cohesion index. The high-density sub-clouds are mapped to potential events, and the confidence of the potential events is evaluated through a confidence ripple feedback mechanism to determine the event information. A topology-temporal two-layer signature algorithm is used to uniquely identify the entity information, the relationships, and the event information to construct a dynamic knowledge graph.

[0041] It's important to note that 3D slicing is an effective method for finely segmenting preprocessed multi-source data streams. By utilizing three dimensions—time windows, spatial grids, and semantic themes—a continuous and complex data stream can be divided into a series of relatively independent and distinctive slice data packets. Specifically, this involves: dividing the continuous multi-source data stream into multiple data segments within a preset time window. These time windows can be of fixed duration, such as per minute or hour, or dynamically adjusted based on data characteristics. Next, within each time window, the data is further divided into multiple spatial grids based on spatial location information. The division of spatial grids can be based on geographic coordinates, grid coding, or other spatial indexing methods to ensure that the data within each grid has similar geographical location characteristics. Finally, for each time window-spatial grid combination, further subdivision is performed based on the semantic theme of the data. Semantic themes can be determined based on the data's source, type, or content, such as traffic flow, traffic accidents, or weather conditions. Through this 3D slicing process, the original multi-source data stream is transformed into a series of independently evolving slice data packets, each containing data information within a specific time, space, and semantic range, facilitating subsequent knowledge extraction and graph construction.

[0042] Emergence value is used to measure the probability and importance of new entities, relationships, or events occurring in a slice data packet. By calculating the emergence value, slice data packets that may contain important new information can be quickly filtered out, providing targeted data sources for subsequent entity recognition and knowledge extraction. For example, in the transportation field, a high emergence value in a slice data packet may indicate that a new traffic event or new traffic participants have occurred within that time period and spatial area. The formula for calculating the emergence value is:

[0043]

[0044] in, For sliced ​​packages Emergence value, For the frequency of new entity appearance, For the new relationship to gain strength, Information entropy reflects the degree of semantic confusion. , , These are weighting coefficients, which can be learned through training with historical event annotations.

[0045] Slice data packets whose emergence values ​​reach a preset emergence threshold are designated as entity incubation candidate sets. These candidate sets are the focus of subsequent entity recognition and knowledge extraction. By setting a reasonable emergence threshold, it can be ensured that only slice data packets that truly contain important new information are included in the candidate set, thereby improving the accuracy and efficiency of entity recognition.

[0046] Based on the characteristic distribution of data packets in the entity incubation candidate set, an adaptive clustering algorithm is used to identify potential entity clusters. The adaptive clustering algorithm can automatically adjust the clustering parameters and strategies according to the actual distribution of the data, thereby more accurately identifying potential entity clusters in the data. For example, in the transportation field, by performing cluster analysis on the data in the entity incubation candidate set, similar traffic participants (such as vehicles of the same type or pedestrians with similar behavioral characteristics) can be grouped into the same potential entity cluster, providing a foundation for subsequent entity information extraction and pattern matching.

[0047] For each latent entity cluster, feature extraction and pattern matching are performed to obtain entity information. Feature extraction extracts representative and distinguishable features from the latent entity cluster to accurately describe and identify the entities. Pattern matching compares the extracted features with known entity patterns to determine the specific type and attributes of the entity. For example, in the transportation field, by performing feature extraction and pattern matching on the latent vehicle entity cluster, information such as vehicle type (e.g., sedan, truck), color, and license plate number can be identified.

[0048] Relationship pulse sniffing is a technique that captures dynamic relationships between entities. By analyzing the interactive behaviors and feature changes of entities in different time, space, and semantic environments, potential relationships between entities can be discovered. For example, in the transportation field, relationship pulse sniffing can discover relationships between vehicles, between vehicles and pedestrians, and between vehicles and traffic facilities, such as following relationships, overtaking relationships, and collision relationships.

[0049] Transient cloud graphs are constructed using entity information and relationships as nodes, and high-density sub-clouds within these transient cloud graphs are identified using a cluster cloud cohesion index. A transient cloud graph is a data structure that visually displays the distribution of entities and relationships. By constructing the transient cloud graph using entity information and relationships as nodes, the relationships and distribution characteristics between different entities can be clearly seen. The cluster cloud cohesion index measures the density of different sub-clouds within the transient cloud graph. By identifying high-density sub-clouds, groups of entities with close relationships can be discovered, providing important clues for subsequent event identification. The formula for cluster cloud cohesion is:

[0050]

[0051] in, Transient graph cloud The cloud cohesion value, Transient graph cloud The set of middle edges, Transient graph cloud The set of nodes in the middle, Transient graph cloud The diameter, which is the length of the longest path in the graph. This is the preset maximum diameter threshold.

[0052] By calculating the cluster cloud cohesion value, the density of different sub-clouds within a transient cloud image can be quantified, thereby identifying high-density sub-clouds. High-density sub-clouds are then mapped to potential events, inferring specific possible events based on the density of entities and their relationships within them. For example, in the transportation field, if a high-density sub-cloud contains multiple vehicles and their collision relationships, it can be mapped to a potential traffic accident event.

[0053] The confidence ripple feedback mechanism is used to assess the confidence of potential events and determine event information. This mechanism comprehensively considers multiple factors to evaluate the confidence of an event. By analyzing relevant entities, relationships, time, and spatial information of the potential event, and combining historical data and domain knowledge, a confidence value for the potential event is determined. Based on the confidence value, the authenticity and importance of the event can be determined, thus obtaining accurate event information. For example, in the transportation field, by mapping high-density subclouds to potential traffic accident events and evaluating them using the confidence ripple feedback mechanism, it is possible to determine whether a traffic accident actually occurred and its severity.

[0054] A topology-temporal dual-layer signature algorithm is employed to uniquely identify entity information, relationships, and event information. Constructing a dynamic knowledge graph is crucial to prevent duplicate entries or version confusion; each entity, relationship, and event must be assigned a globally unique and traceable ID. The topology-temporal dual-layer signature algorithm is an identification algorithm that simultaneously considers the topological structure and temporal characteristics of data. By uniquely identifying entity information, relationships, and event information, each element in the knowledge graph has a unique identity, facilitating subsequent querying, updating, and management. Furthermore, the dynamic knowledge graph can evolve in real-time as new data is added and old data is updated, providing continuous and accurate data support for the adaptive decision-making of domain-specific intelligent agents. For example, in the transportation field, constructing a dynamic knowledge graph allows for real-time monitoring of the operational status and trends of the transportation system, providing strong support for the decision-making of intelligent transportation management systems.

[0055] Step S30: Obtain the current decision task, and search for candidate decision paths on the dynamic knowledge graph based on the current decision task using reinforcement learning and path reasoning algorithms.

[0056] It's important to note that reinforcement learning is a machine learning method that involves an agent interacting with its environment and continuously adjusting its behavioral strategies based on reward signals from the environment to maximize long-term cumulative rewards. Path reasoning algorithms, on the other hand, are based on existing knowledge graph structures, combined with specific rules and logic, to select paths from numerous possibilities that meet specific conditions or objectives.

[0057] This application combines reinforcement learning with path reasoning algorithms to search for candidate decision paths on dynamic knowledge graphs, offering significant advantages. Reinforcement learning enables agents to adaptively explore different decision paths based on constantly changing information in the dynamic knowledge graph, continuously optimizing their search strategies through environmental feedback rewards to find better paths. Path reasoning algorithms, on the other hand, utilize explicit entity, relationship, and event information within the dynamic knowledge graph, based on pre-defined rules and logic, to quickly and accurately filter candidate decision paths that meet the requirements of the current decision task. For example, in the transportation field, the current decision task might be planning the optimal travel route from origin to destination, while considering factors such as real-time traffic conditions, road construction, and traffic accidents. By searching for candidate decision paths on the dynamic knowledge graph using reinforcement learning and path reasoning algorithms, the agent can continuously adjust its search direction based on real-time updated traffic information, avoiding congested sections and construction areas, and finding the shortest and smoothest route as a candidate decision path. This combination fully leverages the adaptive learning capabilities of reinforcement learning and the precise filtering capabilities of path reasoning algorithms, improving the efficiency and accuracy of decision path search and providing high-quality candidate solutions for subsequent decisions.

[0058] Step S40: Use the policy gradient algorithm, combined with the domain reward function, to optimize and rank the candidate decision paths, and generate the optimal decision strategy.

[0059] It should be noted that the policy gradient algorithm is a type of reinforcement learning algorithm that directly optimizes policy parameters. It maximizes long-term cumulative rewards by calculating the gradient of the policy and updating the policy parameters along the gradient direction.

[0060] In this application, a policy gradient algorithm is employed to optimize and rank candidate decision paths. This algorithm adaptively adjusts the priority of each candidate path based on the specific needs of the current decision task and real-time information from the dynamic knowledge graph. The domain reward function is designed according to the characteristics and decision objectives of a specific domain, used to quantify the rewards or benefits brought by different decision paths. For example, in the transportation domain, the domain reward function can comprehensively consider multiple factors such as travel time, route congestion, and road safety to calculate a comprehensive reward value for each candidate decision path. By combining the policy gradient algorithm with the domain reward function, candidate decision paths can be accurately optimized and ranked, prioritizing paths that bring higher rewards and better align with the decision objectives, ultimately generating the optimal decision strategy. This optimal decision strategy fully considers real-time environmental information and domain-specific needs, providing accurate and effective decision guidance for the domain agent and helping it make optimal choices in complex and ever-changing environments. For example, in traffic management scenarios, the optimal decision strategy can guide intelligent transportation systems to rationally allocate traffic flow, alleviate congestion, and improve overall traffic efficiency.

[0061] In one feasible implementation, step S40 may include: defining a domain reward function, which is used to assign a corresponding reward value to each candidate decision path based on the effectiveness, timeliness, resource consumption, and domain-specific constraints of the decision path; calculating the total reward value of each candidate decision path according to the domain reward function, wherein the total reward value is the sum or weighted sum of the reward values ​​of each node on the candidate decision path; using a policy gradient algorithm to perform gradient ascent optimization on the candidate decision paths according to the total reward value, adjusting the action selection probability distribution of each node on the path to obtain an optimized action selection probability distribution; and reordering the candidate decision paths according to the optimized action selection probability distribution, selecting the candidate decision path with the highest total reward value as the optimal decision strategy.

[0062] It should be noted that the domain reward function is used to assign corresponding reward values ​​to each candidate decision path based on its effectiveness, timeliness, resource consumption, and domain-specific constraints. Effectiveness refers to whether the problem can be solved effectively; timeliness refers to the responsiveness of the response; resource consumption refers to the amount of human and material resources used; and domain-specific constraints refer to whether they violate rules. The formula for the domain reward function is:

[0063] in, Decision path Total reward value, For validity score, For timeliness, Total resource consumption The total number of violations. , , , These are preset weighting coefficients, which can be adjusted according to specific application scenarios and decision-making objectives.

[0064] The calculated total reward value comprehensively reflects the performance of candidate decision paths across multiple key dimensions. In the transportation domain, the effectiveness of a decision path might be reflected in whether it can smoothly reach the destination from the origin, while timeliness is closely related to travel time. Resource consumption can consider fuel consumption or electricity consumption, and domain-specific constraints might include whether it passes through a specific area or whether it complies with traffic rules. The total reward value calculated through the domain reward function can accurately quantify the merits of each candidate decision path. For example, in a transportation scenario, if a candidate decision path can quickly reach the destination but violates traffic rules, the total number of violations... This will significantly increase, leading to a decrease in the total reward value. The first path, while taking slightly longer, is fully compliant and consumes fewer resources, and may therefore receive a higher timeliness score. and validity score Achieve a better sorting.

[0065] The policy gradient algorithm optimizes the action selection probability distribution of each node on the path through gradient ascent. For example, in traffic route planning, if a road segment's timeliness score decreases due to an accident, the algorithm will reduce the probability of that road segment being selected while increasing the priority of other feasible paths. Finally, the candidate paths are reordered based on the optimized action selection probability distribution, and the path with the highest total reward value is selected as the optimal decision strategy, ensuring that the decision result both meets domain constraints and maximizes overall benefits.

[0066] Step S50: Feed back the optimal decision strategy to the task initiator to guide the domain agent to make adaptive decisions.

[0067] It's important to note that a domain-specific intelligent agent is an intelligent system deployed within a specific domain environment, capable of autonomously executing tasks and adapting to environmental changes based on input information and decision-making strategies. After feeding back the optimal decision-making strategy to the task initiator, the domain-specific intelligent agent can adjust its own behavior patterns based on that strategy. For example, in traffic management scenarios, after receiving the optimal decision-making strategy, the intelligent transportation system, acting as a domain-specific intelligent agent, will dynamically adjust traffic light timings, optimize lane allocation schemes, and even send navigation suggestions to vehicles through vehicle-to-infrastructure (V2I) technology. This feedback mechanism ensures the integrity of the decision-making closed loop—the complete chain from environmental perception to strategy generation and then to execution feedback enables the intelligent agent to continuously adapt to complex and ever-changing real-world scenarios.

[0068] This embodiment provides a domain agent adaptive decision-making method based on a dynamic knowledge graph. It collects raw multi-source data from multiple heterogeneous data sources in real time, and performs three-dimensional indexing preprocessing (time-space-semantic) on the raw multi-source data at the edge, obtaining preprocessed multi-source data. Based on the preprocessed multi-source data, incremental knowledge extraction technology is used to dynamically identify and extract entity information, relationships, and event information, and a dynamic knowledge graph is constructed based on the entity information, relationships, and event information. The current decision task is obtained, and based on the current decision task, reinforcement learning and path reasoning algorithms are used to search for candidate decision paths on the dynamic knowledge graph. A policy gradient algorithm, combined with a domain reward function, is used to optimize and rank the candidate decision paths to generate the optimal decision strategy. The optimal decision strategy is fed back to the task initiator to guide the domain agent in completing adaptive decision-making. Through the above method, by collecting and preprocessing multi-source heterogeneous data in real time, constructing a dynamic knowledge graph, and performing adaptive learning and decision-making based on the graph, the real-time performance and accuracy of the domain agent's decision-making can be effectively improved.

[0069] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 Step S30 includes steps S301 to S303:

[0070] Step S301: Perform multi-scale compression and subgraph partitioning on the dynamic knowledge graph to generate cascaded subgraphs adapted to different decision-making tasks.

[0071] It's important to note that multi-scale compression aims to retain key information in the knowledge graph while reducing data redundancy and improving subsequent processing efficiency. For example, in a traffic dynamics knowledge graph, compression can be performed by regional scope (e.g., city-level, district-level) or time granularity (e.g., hourly, minute-level) to remove low-frequency entities and redundant relationships. Subgraph partitioning breaks down the complete graph into multiple logical subsets based on decision-making task requirements. For instance, for a morning rush hour congestion mitigation task, subgraphs related to main roads and key intersections can be extracted, while excluding branch road information irrelevant to the current task. Cascaded subgraphs organize subgraphs of different granularities through a hierarchical structure, forming a hierarchical relationship of global overview to local details. This supports both macro-level decision analysis and allows for focusing on specific scenarios. For example, in a logistics scheduling scenario, the top-level subgraph displays the distribution of the national warehousing network, the middle-level subgraph presents the relationships between regional distribution centers, and the bottom-level subgraph details the storage status of goods in specific warehouses. Different levels of subgraphs are dynamically linked through entity IDs and relationship types. This structure enables the system to adaptively call subgraphs at different levels based on the complexity of the decision-making task, reducing computational overhead while ensuring decision-making accuracy.

[0072] In one feasible implementation, step S301 may include: performing preliminary compression on the dynamic knowledge graph using a graph clustering algorithm to generate multiple subgraph modules; performing recursive secondary compression on the subgraph modules by setting different compression granularity parameters to form a multi-layered cascaded subgraph structure; performing three-dimensional density pulse scanning on the nodes in the dynamic knowledge graph to generate a density pulse map; determining multi-scale segmentation boundaries based on the density peaks and valleys in the density pulse map; and segmenting and adjusting the multi-layered cascaded subgraph structure according to the multi-scale segmentation boundaries to generate cascaded subgraphs adapted to different decision-making tasks.

[0073] It should be noted that graph clustering algorithms divide the knowledge graph into tightly structured sub-modules by analyzing the connections between nodes. For example, in a traffic graph, road nodes in the same area can be clustered into independent modules. Recursive secondary compression optimizes the initially compressed sub-modules by dynamically adjusting compression granularity parameters (such as regional range thresholds and time window lengths), forming a multi-level subgraph structure that includes city-level backbone networks and district-level branch road networks.

[0074] Three-dimensional density pulse scanning generates a pulse map containing density peaks (such as transportation hubs) and density valleys (such as remote areas) by statistically analyzing the correlation density of nodes in time, space, and semantic dimensions. For example, it can identify high-frequency scheduling warehouse nodes in a logistics map.

[0075] Multi-scale segmentation boundaries are determined based on the extreme points of the density pulse map. For example, the boundaries of regions with abrupt density changes are used as the basis for subgraph division, ensuring that cascaded subgraphs can cover key decision-making elements while avoiding interference from irrelevant information. The final generated cascaded subgraphs are dynamically linked through entity IDs and relationship types. For example, in an emergency management scenario, the top-level subgraph displays the province-wide disaster monitoring network, the middle-level subgraphs focus on the infrastructure status of disaster-stricken areas, and the bottom-level subgraphs detail the material needs of specific rescue points. Subgraphs of different levels can be flexibly combined and invoked according to task requirements.

[0076] Step S302: Obtain the current decision task, perform a formal transformation on the current decision task, and generate a graph query.

[0077] It should be noted that the purpose of formal transformation is to convert the decision task described in natural language into a query language or query pattern that the knowledge graph can understand, so as to accurately locate relevant information in the dynamic knowledge graph. That is, to convert the current decision task T into a sequence of logical predicates, as shown in the following formula:

[0078] in, The graph query generated for the current decision task T. The number of logical predicates, Indicates the first Logical predicates, such as "located in", "associated with", "belonging to", etc. and These are the subject and object entities of the predicate, respectively.

[0079] In traffic management scenarios, if the decision task is "to query the average vehicle speed on a main road during the morning rush hour," the transformed query might generate a sequence containing predicates such as "time = morning rush hour," "road type = main road," and "indicator = average vehicle speed." This transformation converts fuzzy natural language requirements into structured graph queries, enabling efficient retrieval of relevant entities and relationships within a dynamic knowledge graph.

[0080] In practical implementation, natural language processing techniques, such as named entity recognition and relation extraction, can be used to parse the task description, extract key entities (such as time, location, and indicator type) and relation types, and then generate a logical sequence that conforms to the query syntax of a knowledge graph based on a predefined predicate template. This transformation mechanism supports flexible processing of cross-domain tasks. Through formal transformation, it can shield the language differences between tasks in different domains and uniformly transform them into a logical structure that can be processed by a knowledge graph, thereby improving the compatibility of cross-domain decision-making.

[0081] Step S303: Using reinforcement learning and path reasoning algorithms, search for candidate decision paths on the cascaded subgraphs based on the graph query.

[0082] It's important to note that reinforcement learning continuously optimizes decision-making strategies through the interaction between the agent and its environment. Its core lies in balancing exploration and utilization—trying new paths to discover potentially better solutions while leveraging existing experience to improve search efficiency. In cascaded subgraph search, the agent starts with a graph query, selects actions based on the current subgraph level (such as switching to adjacent subgraphs or traversing along edges with specific relationships), and evaluates the potential benefits of these actions using path reasoning algorithms. For example, in a logistics scheduling scenario, if the query is "the shortest time-sensitive path from regional distribution center A to customer B," the agent might first locate the regions of A and B in the top-level subgraph, then switch to the middle-level subgraph to analyze the relationships between distribution centers and warehouses within the region, and finally plan the specific transportation route in the bottom-level subgraph.

[0083] The path reasoning algorithm combines graph structure features (such as edge weights and node attributes) and domain constraints (such as vehicle load and time windows) to assign multi-dimensional scores, including timeliness, cost, and risk, to candidate paths through a dynamic weight adjustment mechanism. For example, if a path's timeliness score drops due to traffic control, the algorithm will lower its priority while improving the ranking of alternative routes. Reinforcement learning provides feedback on decision quality through a reward function (such as the sum of path scores), driving the agent to gradually converge to the optimal path. This mechanism enables the system to adaptively handle query tasks of varying complexity—simple tasks can be performed locally in the underlying subgraph, while complex tasks achieve global optimization through hierarchical jumps, thereby significantly reducing computational overhead while maintaining decision accuracy.

[0084] In one feasible implementation, step S303 may include: converting the graph query into a state representation in a reinforcement learning environment, wherein the state representation includes the current query node, visited paths, and remaining query requirements; defining an action space in the reinforcement learning environment, wherein the action space includes all feasible edges originating from the current query node and their corresponding jump target nodes; generating an action selection strategy using a policy network based on the state representation and the action space, wherein the action selection strategy is used to select the optimal action at each decision step to maximize long-term rewards; updating the state representation according to the action selection strategy to obtain an updated state representation; exploring paths on the cascaded subgraph based on the updated state representation and the action selection strategy, updating the remaining query requirements information after each action jump; stopping path exploration when the remaining query requirements information meets a preset termination condition or reaches the maximum number of exploration steps, and using the set of explored paths as candidate decision paths.

[0085] It's important to note that transforming graph queries into state representations within a reinforcement learning environment is a crucial step in combining reinforcement learning with knowledge graphs. The core of this approach lies in converting abstract graph query requirements into a perceptible and actionable environmental state for the agent. This state representation must include three aspects: First, the current query node serves as the starting point for path searching. For example, in logistics queries, if the user's request is "the transportation route from warehouse A to customer B," then warehouse A is the initial query node. Second, the visited paths record the nodes and edges traversed by the agent during the search process, used to avoid redundant exploration and support path backtracking. For instance, in a transportation network, if the agent has traversed "main road 1 → branch road 2," this path will be recorded in the state. Third, the remaining query requirements reflect unmet query conditions. For example, in time-sensitive tasks, remaining requirements might include "remaining available time" or "uncovered key nodes." By encoding these three types of information into vector or tensor forms, the state representation can fully characterize the agent's search context within the graph, providing a basis for subsequent action selection.

[0086] Understandably, the action space in a reinforcement learning environment defines all legal operations an agent can take during the search process, and its design must balance the connectivity of the graph structure with domain constraints. In a cascading subgraph scenario, the action space includes two types of actions: first, traversing laterally along the edges within the current subgraph; for example, in a logistics subgraph, the agent can move from "Warehouse A" to "Transfer Station B" along the "Transportation Relationship"; second, switching to adjacent subgraphs through preset jump rules; for example, when the query involves cross-regional transportation, the agent can jump from the "City-level Subgraph" to the "Region-level Subgraph" to continue the search. The dynamism of the action space is reflected in its automatic adjustment based on the current subgraph level and remaining query requirements—lower-level subgraphs may only contain local transportation edges, while top-level subgraphs provide cross-regional jump interfaces. This design allows the agent to focus on local details while also grasping global relationships.

[0087] As a core component of reinforcement learning, the policy network's role is to generate an action selection probability distribution based on the current state representation, guiding the agent to choose the optimal action at each decision step. In its implementation, the policy network can employ a deep neural network structure. The input is the encoded state representation, such as query node embeddings, path history embeddings, and remaining demand embeddings, while the output is the Q-value or probability of each action in the action space. For example, in a traffic management scenario, if the state representation shows the current location is "main road 1" and the remaining demand is "reach intersection X within 10 minutes," the policy network might assign a high probability to "go straight to intersection X" and a low probability to "take a detour." This action selection mechanism based on state context allows the agent to dynamically adapt to the complexity of different query tasks.

[0088] The optimization objective of the action selection strategy is to maximize long-term rewards, where the reward function needs to comprehensively consider multiple dimensions such as timeliness, cost, and risk. The core of state updates is to encode the environmental feedback after the action is executed into the state representation to support reasoning for the next decision step.

[0089] Specifically, when the agent performs the action of "jumping from node A to node B", the state update includes three aspects: First, updating the current queried node to B; second, adding "A→B" to the visited path; and third, adjusting the remaining query requirements based on the action result (such as reducing the remaining time or increasing the covered nodes). For example, in an emergency management scenario, if the initial query is "deploy 5 fire trucks to the disaster area within 30 minutes", after the agent completes the action of "deploying 2 vehicles from warehouse 1", the remaining requirement will be updated to "deploy 3 vehicles to the disaster area within 28 minutes". This dynamic state update mechanism ensures that the agent can continuously perceive environmental changes and make adaptive decisions.

[0090] The termination conditions for path exploration need to balance efficiency and completeness, typically including two types of triggering rules: First, the remaining query requirements meet a preset threshold, such as remaining time reaching zero or all key nodes being covered; second, the maximum number of exploration steps is reached (to prevent infinite loops). For example, in a traffic query, if the user's requirement is "to reach the airport within 20 minutes," if the agent's planned path has an estimated arrival time of 19 minutes, the exploration will terminate and return to the current path even if the maximum number of steps has not been reached. Conversely, if the exploration steps exceed 100 and the requirement is still not met, the exploration will be forcibly terminated and the optimal candidate path will be returned. This termination mechanism ensures timely decision-making while avoiding computational waste caused by over-exploration.

[0091] The final set of candidate decision paths is the result of joint optimization of reinforcement learning and path reasoning. Its quality depends on the accuracy of the state representation, the generalization ability of the policy network, and the rationality of the reward function. This set of paths will be further sorted and optimized by the policy gradient algorithm, and finally output the optimal decision policy that takes into account both efficiency and feasibility, thereby guiding the domain agent to complete adaptive decision-making.

[0092] In this embodiment, by using a dynamic generation mechanism of multi-scale segmentation boundaries and cascaded subgraphs, combined with a reinforcement learning-driven path search strategy, decision paths are efficiently extracted from complex graph structures, effectively improving the adaptability and accuracy of the domain agent in handling diverse decision-making tasks.

[0093] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the domain intelligent agent adaptive decision-making method based on dynamic knowledge graphs in this application. Any simple transformations based on this technical concept are within the protection scope of this application.

[0094] This application also provides a domain-specific intelligent agent adaptive decision-making device based on dynamic knowledge graphs. Please refer to [link / reference]. Figure 3 The domain-specific adaptive decision-making device based on dynamic knowledge graphs includes: The preprocessing module 10 is used to collect raw multi-source data from multiple heterogeneous data sources in real time, and perform three-dimensional index preprocessing of the raw multi-source data in terms of time, space and semantics on the edge side to obtain preprocessed multi-source data.

[0095] The construction module 20 is used to dynamically identify and extract entity information, relationship and event information based on the preprocessed multi-source data using incremental knowledge extraction technology, and to construct a dynamic knowledge graph based on the entity information, relationship and event information.

[0096] Search module 30 is used to obtain the current decision task and search for candidate decision paths on the dynamic knowledge graph based on the current decision task using reinforcement learning and path reasoning algorithms.

[0097] The sorting module 40 is used to optimize and sort the candidate decision paths by using a policy gradient algorithm combined with a domain reward function, and generate the optimal decision strategy.

[0098] Feedback module 50 is used to feed back the optimal decision strategy to the task initiator to guide the domain agent to complete adaptive decision-making.

[0099] The domain agent adaptive decision-making device based on dynamic knowledge graphs provided in this application, employing the domain agent adaptive decision-making method based on dynamic knowledge graphs in the above embodiments, can solve the technical problems of low decision-making accuracy, poor real-time performance, and weak adaptability of domain agents. Compared with the prior art, the beneficial effects of the domain agent adaptive decision-making device based on dynamic knowledge graphs provided in this application are the same as those of the domain agent adaptive decision-making method based on dynamic knowledge graphs provided in the above embodiments, and other technical features in the domain agent adaptive decision-making device based on dynamic knowledge graphs are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0100] This application provides a domain intelligent agent adaptive decision-making device based on dynamic knowledge graph. The domain intelligent agent adaptive decision-making device based on dynamic knowledge graph includes: at least one processor; and a memory communicatively connected to at least one processor; wherein the memory stores instructions executable by at least one processor, and the instructions are executed by at least one processor to enable at least one processor to execute the domain intelligent agent adaptive decision-making method based on dynamic knowledge graph in the above embodiment 1.

[0101] The following is for reference. Figure 4 This document illustrates a structural schematic diagram of a domain intelligent agent adaptive decision-making device based on a dynamic knowledge graph, suitable for implementing embodiments of this application. The domain intelligent agent adaptive decision-making device based on a dynamic knowledge graph in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 4 The domain intelligent agent adaptive decision-making device based on dynamic knowledge graph shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0102] like Figure 4As shown, the domain-specific adaptive decision-making device based on a dynamic knowledge graph may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in ROM (Read Only Memory) 1002 or a program loaded from storage device 1003 into RAM (Random Access Memory) 1004. RAM 1004 also stores various programs and data required for the operation of the domain-specific adaptive decision-making device based on the dynamic knowledge graph. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. Input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, LCDs (Liquid Crystal Displays), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the domain agent adaptive decision-making device based on dynamic knowledge graphs to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows a domain agent adaptive decision-making device based on dynamic knowledge graphs with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0103] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0104] The domain agent adaptive decision-making device based on dynamic knowledge graphs provided in this application, employing the domain agent adaptive decision-making method based on dynamic knowledge graphs in the above embodiments, can solve the technical problems of low decision-making accuracy, poor real-time performance, and weak adaptability of domain agents. Compared with the prior art, the beneficial effects of the domain agent adaptive decision-making device based on dynamic knowledge graphs provided in this application are the same as those of the domain agent adaptive decision-making method based on dynamic knowledge graphs provided in the above embodiments, and other technical features in this domain agent adaptive decision-making device based on dynamic knowledge graphs are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0105] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0106] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0107] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the domain intelligent agent adaptive decision-making method based on dynamic knowledge graph in the above embodiments.

[0108] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, RAM (Random Access Memory), ROM (Read Only Memory), EPROM (Erasable Programmable Read Only Memory or Flash Memory), optical fibers, CD-ROM (CD-Read Only Memory), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0109] The aforementioned computer-readable storage medium may be included in a domain intelligent agent adaptive decision-making device based on dynamic knowledge graphs; or it may exist independently and not be assembled into a domain intelligent agent adaptive decision-making device based on dynamic knowledge graphs.

[0110] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by a domain-based intelligent agent adaptive decision-making device based on a dynamic knowledge graph, the device performs the following actions: real-time acquisition of raw multi-source data from multiple heterogeneous data sources; performs three-dimensional indexing preprocessing (time-space-semantic) on the raw multi-source data at the edge to obtain preprocessed multi-source data; uses incremental knowledge extraction technology based on the preprocessed multi-source data to dynamically identify and extract entity information, relationships, and event information, and constructs a dynamic knowledge graph based on the entity information, relationships, and event information; acquires the current decision task; searches for candidate decision paths on the dynamic knowledge graph based on the current decision task using reinforcement learning and path reasoning algorithms; optimizes and ranks the candidate decision paths using a policy gradient algorithm combined with a domain reward function to generate an optimal decision strategy; and feeds back the optimal decision strategy to the task initiator to guide the domain agent in completing adaptive decision-making.

[0111] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including LAN (Local Area Network) or WAN (Wide Area Network)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0112] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0113] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0114] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the aforementioned adaptive decision-making method for domain intelligent agents based on dynamic knowledge graphs. This addresses the technical problems of low decision-making accuracy, poor real-time performance, and weak adaptability of domain intelligent agents. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the adaptive decision-making method for domain intelligent agents based on dynamic knowledge graphs provided in the above embodiments, and will not be elaborated upon here.

[0115] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the adaptive decision-making method for a domain intelligent agent based on a dynamic knowledge graph as described above.

[0116] The computer program product provided in this application can solve the technical problems of low decision-making accuracy, poor real-time performance, and weak adaptability of domain intelligent agents. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the adaptive decision-making method for domain intelligent agents based on dynamic knowledge graphs provided in the above embodiments, and will not be repeated here.

[0117] The above are only some embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A domain-specific adaptive decision-making method for intelligent agents based on dynamic knowledge graphs, characterized in that, The method includes: Raw multi-source data from multiple heterogeneous data sources is collected in real time, and the raw multi-source data is preprocessed with a three-dimensional index of time, space and semantics at the edge to obtain preprocessed multi-source data. Based on the preprocessed multi-source data, incremental knowledge extraction technology is used to dynamically identify and extract entity information, relationships and event information, and a dynamic knowledge graph is constructed based on the entity information, relationships and event information. Obtain the current decision task, and search for candidate decision paths on the dynamic knowledge graph based on the current decision task using reinforcement learning and path reasoning algorithms; The candidate decision paths are optimized and ranked using a policy gradient algorithm combined with a domain reward function to generate the optimal decision strategy. The optimal decision-making strategy is fed back to the task initiator to guide the domain agent in making adaptive decisions.

2. The method as described in claim 1, characterized in that, The process involves real-time acquisition of raw multi-source data from multiple heterogeneous data sources, followed by edge-side preprocessing of the raw multi-source data using a three-dimensional temporal-spatial-semantic index, resulting in preprocessed multi-source data, including: A distributed data acquisition framework is used to perform real-time synchronous acquisition of multi-source heterogeneous data to obtain raw multi-source data. A spatiotemporal semantic coding model is deployed through edge computing nodes, and the original multi-source data is indexed and encoded in terms of time, space and semantic dimensions through the spatiotemporal semantic coding model to construct a three-dimensional index structure that includes temporal features, spatial distribution and semantic association. Based on the three-dimensional index structure, the original multi-source data is retrieved and located to extract multi-source data fragments related to the current decision-making task. A lightweight data cleaning rule engine is used to perform outlier filtering, missing value filling and unit unification preprocessing on the multi-source data fragments at the edge, generating preprocessed multi-source data.

3. The method as described in claim 1, characterized in that, The preprocessed multi-source data employs incremental knowledge extraction technology to dynamically identify and extract entity information, relationships, and event information, and constructs a dynamic knowledge graph based on the entity information, relationships, and event information, including: The preprocessed multi-source data stream is sliced ​​in three dimensions according to time window, spatial grid, and semantic topic to form independently evolving slice data packets; Emergence is calculated for each slice data packet to obtain the emergence value of each slice data packet, wherein the emergence value is used to measure the probability and importance of the occurrence of new entities, new relationships or new events in the slice data packet; Slice data packets whose emergence values ​​reach a preset emergence threshold are used as entity incubation candidate sets; Based on the characteristic distribution of data packets in the entity incubation candidate set, an adaptive clustering algorithm is used to identify potential entity clusters, and feature extraction and pattern matching are performed on each potential entity cluster to obtain entity information; The entity information is subjected to relationship pulse sniffing to identify the relationships between entities; The entity information and the relationship are used as nodes to form a transient cloud, and the high-density sub-clouds in the transient cloud are identified by the cloud cluster cohesion index. The high-density sub-cloud is mapped to potential events, and the confidence of the potential events is evaluated through a confidence ripple feedback mechanism to determine the event information; A topology-temporal dual-layer signature algorithm is used to uniquely identify the entity information, the relationship, and the event information to construct a dynamic knowledge graph.

4. The method as described in claim 1, characterized in that, The step of obtaining the current decision task and searching for candidate decision paths on the dynamic knowledge graph based on the current decision task using reinforcement learning and path reasoning algorithms includes: The dynamic knowledge graph is subjected to multi-scale compression and subgraph partitioning to generate cascaded subgraphs adapted to different decision-making tasks; Obtain the current decision task, perform a formal transformation on the current decision task, and generate a graph query; Reinforcement learning and path reasoning algorithms are used to search for candidate decision paths on the cascaded subgraphs based on the graph query.

5. The method as described in claim 4, characterized in that, The process of performing multi-scale compression and subgraph partitioning on the dynamic knowledge graph to generate cascaded subgraphs adapted to different decision-making tasks includes: The dynamic knowledge graph is initially compressed using a graph clustering algorithm to generate multiple subgraph modules; By setting different compression granularity parameters, the subgraph module is recursively compressed twice to form a multi-layered cascaded subgraph structure; Perform a three-dimensional density pulse scan on the nodes in the dynamic knowledge graph to generate a density pulse map; Based on the density peak and density valley values ​​in the density pulse map, the multi-scale segmentation boundary is determined. The multi-level cascaded subgraph structure is segmented and adjusted according to the multi-scale segmentation boundary to generate cascaded subgraphs adapted to different decision-making tasks.

6. The method as described in claim 4, characterized in that, The step of using reinforcement learning and path reasoning algorithms to search for candidate decision paths on the cascaded subgraphs based on the graph query includes: The graph query is transformed into a state representation in a reinforcement learning environment, wherein the state representation includes information on the current query node, visited paths, and remaining query requirements. In a reinforcement learning environment, an action space is defined, which includes all feasible edges originating from the current query node and their corresponding target nodes. Based on state representation and action space, an action selection policy is generated using a policy network, wherein the action selection policy is used to select the optimal action at each decision step to maximize long-term reward. The state representation is updated according to the action selection strategy to obtain the updated state representation; Path exploration is performed on the cascaded subgraph based on the updated state representation and action selection strategy. After each action jump is completed, the remaining query requirement information is updated. When the remaining query requirements meet the preset termination conditions or the maximum number of exploration steps is reached, path exploration stops, and the set of paths already explored is used as candidate decision paths.

7. The method as described in claim 1, characterized in that, The step involves employing a policy gradient algorithm, combined with a domain reward function, to optimize and rank the candidate decision paths, generating the optimal decision strategy, including: Define a domain reward function, which is used to assign a corresponding reward value to each candidate decision path based on the effectiveness, timeliness, resource consumption, and domain-specific constraints of the decision path; The total reward value for each candidate decision path is calculated based on the domain reward function, wherein the total reward value is the sum or weighted sum of the reward values ​​of each node on the candidate decision path; The policy gradient algorithm is used to optimize the candidate decision path by gradient ascent based on the total reward value, and the action selection probability distribution of each node on the path is adjusted to obtain the optimized action selection probability distribution. Based on the optimized action selection probability distribution, the candidate decision paths are reordered, and the candidate decision path with the highest total reward value is selected as the optimal decision strategy.

8. A domain-specific intelligent agent adaptive decision-making device based on dynamic knowledge graphs, characterized in that, The device includes: The preprocessing module is used to collect raw multi-source data from multiple heterogeneous data sources in real time, and perform three-dimensional index preprocessing of the raw multi-source data in terms of time, space and semantics on the edge side to obtain preprocessed multi-source data. The construction module is used to dynamically identify and extract entity information, relationship and event information based on the preprocessed multi-source data using incremental knowledge extraction technology, and to construct a dynamic knowledge graph based on the entity information, relationship and event information; The search module is used to obtain the current decision task and, based on the current decision task, search for candidate decision paths on the dynamic knowledge graph using reinforcement learning and path reasoning algorithms. The ranking module is used to optimize and rank the candidate decision paths using a policy gradient algorithm combined with a domain reward function, thereby generating the optimal decision strategy. The feedback module is used to feed back the optimal decision-making strategy to the task initiator to guide the domain agent to complete adaptive decision-making.

9. A domain-specific intelligent agent adaptive decision-making device based on dynamic knowledge graphs, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the domain agent adaptive decision-making method based on dynamic knowledge graphs as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a domain agent adaptive decision-making program based on a dynamic knowledge graph. When the domain agent adaptive decision-making program based on the dynamic knowledge graph is executed by the processor, it implements the domain agent adaptive decision-making method based on the dynamic knowledge graph as described in any one of claims 1 to 7.

Images