Task-driven command model dynamic arrangement system
By building a command model and knowledge graph, combining semantic analysis and task decomposition, a task-driven command model dynamic orchestration system is realized, solving the problem of selecting massive model resources, improving decision-making efficiency and scientificity, and shortening the closed-loop time of the kill chain.
Patent Information
- Application Number
- CN202411064747.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-05
- Publication Date
- 2024-12-10
AI Technical Summary
In the future battlefield, a large number of command model resources make it difficult for commanders or staff to quickly and accurately select and combine appropriate models to assist decision-making, resulting in a scientific decision-making and an extended closed-loop time of the kill chain.
By building command models, command model knowledge graphs and task handling domain knowledge, semantic analysis, task decomposition and model orchestration methods are used to realize a task-driven command model dynamic orchestration system, and support commanders or staff members to quickly match and combine applicable models. resource.
It improves the scientificity and rapidity of command decisions, shortens the closed-loop time of the kill chain, and reduces the time and cost of decision making.
Smart Images
Figure CN119106846A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of command post command technology, and in particular to a task-driven command model dynamic arrangement system. Background Art
[0002] The future battlefield space will expand to high dimensions of land, sea, air, space, network, and electricity. The explosive growth of battlefield information, the ever-changing battlefield situation, and the significantly accelerated rhythm of war have put forward higher requirements for the timeliness, accuracy, and sensitivity of command decisions. The fast, accurate, and precise command decisions on complex battlefields have become the decisive factors for winning the initiative in war. As a carrier and important form of human knowledge, the model is the basis for realizing artificial intelligence and can greatly enhance human cognition, prediction, decision-making, and control capabilities. The command model is not only the basic support for realizing military intelligence, but also an important tool for realizing the agile development of software systems in the field of command and control. It will become an important strategic asset in the era of digital intelligence.
[0003] In the future, as the construction of command models continues to advance, model resources will become more abundant, and it is expected to provide more comprehensive intellectual support for commanders or staff. However, the massive command model resources also increase the search space for commanders or staff; at the same time, combat needs are full of uncertainty. Requiring everyone to be familiar with all the models that may be used in the future requires very high training costs. The battlefield of the future is changing rapidly, and opportunities are fleeting. Whether it can support commanders or staff to quickly and accurately select the most suitable command model for the current task from massive model resources, and combine them for decision-making assistance, has become the key to improving the scientific nature of decision-making and shortening the closed-loop time of the kill chain.
[0004] In recent years, the vigorous development of technologies such as large language models and knowledge graphs has brought opportunities and challenges that cannot be ignored to many industries, and is also expected to bring changes to the command and control field. This patent proposes an agile orchestration method for command model resources to provide support for improving command and control capabilities. Summary of the invention
[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a task-driven command model dynamic orchestration system, in order to realize the efficient task-oriented dynamic orchestration of command model resources, it is necessary to construct three types of resources: command model, command model knowledge graph and task handling domain knowledge, and use them through the three main links of semantic analysis, task decomposition and model orchestration; In terms of resource construction, three types of resources are mainly constructed, namely command model, command model knowledge graph and task handling domain knowledge.
[0006] Command model: As a software resource, the command model can be constructed by referring to the software engineering method and following the process of demand generation, model design, research and development, verification and validation, and iterative improvement. The model that has completed verification and validation has relatively high credibility and interpretability and can be included in the model resource pool. The model resource pool can be logically integrated and physically distributed.
[0007] Command model knowledge graph: To achieve the discovery and call of model resources, it is necessary to describe the model digitally, which can be achieved by building a command model knowledge graph. The command model knowledge graph is based on ontology and covers important information such as model application scenarios, supported command activities, operating software and hardware environment, supporting data, and relationships between models, to achieve a multi-dimensional description of the model.
[0008] Task handling domain knowledge: To support matching the corresponding model resources according to the analysis and computing needs described by the user, it is necessary to establish task handling domain knowledge. Based on workflow flexible modeling, hierarchical task network and other technologies, describe the common business scenarios of users and the corresponding command workflows, and establish corresponding command activity decomposition rules.
[0009] In terms of organizational application, it includes three main links, namely semantic analysis, task decomposition and model orchestration, and its target users are commanders or staff members.
[0010] Semantic parsing: For analytical computing requirements described by users in natural language, technologies such as large language models can be used to accurately convert them into unambiguous task instructions that can be understood by computers through semantic parsing.
[0011] Task decomposition: Since the user's task instructions may be relatively macroscopic, they need to be decomposed into finer granularity to facilitate mapping them into specific models. Hierarchical task network technology can be used to decompose the tasks layer by layer into granularity that matches the command model resources, supported by the knowledge of task handling domain.
[0012] Model orchestration: For tasks that have been decomposed, the corresponding model resources are matched with the support of the command model knowledge graph. Existing models may fully realize, partially realize, or completely fail to realize the analysis and computing requirements corresponding to a certain atomic task. Model orchestration is to optimize the model or model combination with the support of the command model knowledge graph, and to match the model resources and analysis and computing tasks as much as possible. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] Figure 1 A schematic diagram of agile arrangement of resources of the command model of the present invention; Figure 2 It is a schematic diagram of the model type of the present invention; Figure 3 A schematic diagram of the command activities supported by the model of the present invention; Figure 4 This is a schematic diagram of the use scenario of the model of the present invention; Figure 5 It is a user schematic diagram of the model of the present invention; Figure 6 This is a schematic diagram of data related to the model of the present invention; Figure 7 This is a schematic diagram of the model operation information of the present invention. DETAILED DESCRIPTION
[0014] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The embodiments of the present invention are provided for the purpose of illustration and description, and are not intended to be exhaustive or to limit the present invention to the disclosed forms. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiments are selected and described in order to better illustrate the principles and practical applications of the present invention, and to enable those of ordinary skill in the art to understand the present invention and thereby design various embodiments with various modifications suitable for specific uses.
[0015] The first embodiment, as Figure 1 to Figure 2 As shown, the present invention provides a technical solution: a task-driven command model dynamic orchestration system, which consists of the following steps: Step 1: Construction of command model knowledge graph; The command model ontology abstracts the concept hierarchy within the domain and defines the relevant attributes of each concept and the relationship between concepts at the semantic level. It is the basis for building the model resource knowledge graph.
[0016] In order to facilitate knowledge sharing with related fields, a common core ontology can be used as the upper ontology of the command model ontology, and its definitions of basic concepts such as entities, events, time, and space can be used to facilitate future integration with related field ontologies. On this basis, targeted expansion is carried out based on the characteristics of the command model itself. The combat command model is described through model types, supported command activities, applicable scenarios, model users, model-related data, model operation information, and relationships between models, such as Figure 2 shown.
[0017] On this basis, based on the model resource supporting documentation and natural language processing technology, relevant information is automatically extracted from structured and semi-structured data to build and continuously expand the command model knowledge graph. Business personnel mainly complete information verification and make necessary manual adjustments.
[0018] Task semantic parsing based on large language model; Parsing the computing requirements described by commanders or staff in natural language into structured semantics is the basis for recommending a suitable command model. Language expression is diverse, and different sentence patterns may express the same semantics; language expression is also domain-specific, and the same semantics may have different target structures in different scenarios. Drawing on the successful application experience of pre-trained large models such as BERT and GPT, the "pre-trained large model + task fine-tuning" model is adopted to achieve task semantic parsing. A fully pre-trained general model combined with supervised instruction fine-tuning and reinforcement learning based on human feedback provides a solution for computers to understand the semantics of computing requirements proposed by commanders or staff. Considering that the large language model is under great pressure in a single human-computer dialogue mode and lacks deep interaction with the environment, the application of multimodal embodied large models can be explored.
[0019] Step 2: Flexible modeling of command and control workflow; The command process is a specific process consisting of a series of relatively clear, orderly, and connected stages or basic links. It is the basic procedure for commanders to engage in command activities and can reflect the laws of combat command activities under different command modes. Diversified combat tasks require flexible reconstruction of the command process. Taking into account the requirements of node flexibility and process flexibility, the flexible modeling of the command and control process is completed based on the Unified Modeling Language (UML) extension mechanism. Among them, node flexibility refers to the process of skipping certain task nodes according to actual conditions during the process operation to improve the efficiency of process execution; process flexibility refers to the dynamic selection of corresponding branches according to the current state and operating conditions during the process operation.
[0020] Step 3: Automatically decompose tasks based on hierarchical task networks; For the commander's task requirements to complete semantic analysis, it is necessary to perform detailed decomposition of tasks. As a kind of planning technology, the Hierarchical Task Network (HTN) can decompose complex and abstract tasks into executable atomic tasks step by step by recursively decomposing and resolving conflicts. This idea of layer-by-layer decomposition and refinement is very consistent with the way commanders analyze combat problems. Hierarchical task network planning usually needs to be based on domain knowledge. For combat task decomposition, we can first build a combat task list, describe combat tasks through unified military terms, clarify the connotation, extension, related task objectives and task indicators of combat tasks, and use this as a common standard and common language for the decomposition of various tasks at all levels. Based on the combat task list, task decomposition can be performed based on the task hierarchy network. How to efficiently process dynamic and uncertain tasks in a distributed manner in a collaborative manner when domain knowledge is not complete is still a problem that needs to be studied.
[0021] Step 4: Command model matching recommendation based on hybrid search; Based on the decomposed analytical computing task requirements, a list of relevant command model resources is generated, similar to the page ranking of relevant results generated by search engines. The effectiveness of matching recommendations is crucial to reducing the time of command decision-making. A hybrid search ranking mode that integrates multiple factors is used to match and recommend command models. The goal is to rank models with stronger relevance in the front, so that users do not have to waste time viewing irrelevant models. First, based on combat applications and model operation requirements, constraint matching rules are constructed and applied. Constraint matching rules refer to restrictive conditions that must be met, such as the software and hardware environment in which the model runs, the input and output relationship between models, etc., which are used to preliminarily screen models to exclude models that obviously do not meet the conditions and reduce the size of the matching model set. Secondly, the models after preliminary screening are used as the matching target set, and optimization matching rules are constructed and applied. The optimization matching rules combine relevance and popularity. The relevance is mainly based on the semantic feature annotation of the command model, and the semantic feature vector of each command model is constructed to achieve the measurement of relevance; the popularity tends to rank models with more queries / calls at the top of the recommendation list. Finally, the weight between relevance and popularity is continuously adjusted according to user feedback. The effectiveness of the recommendation result ranking can be measured in explicit or implicit ways. The explicit way is for users to directly rate the recommendation results, and the implicit way is to count the usage rate of the recommendation results. By establishing a matching recommendation mechanism that integrates multiple factors and continuously improves it, the efficiency and accuracy of the command model matching recommendation can be improved.
[0022] Step 5: Adaptive model service; In order to better support the dynamic task-oriented orchestration of command models, it is necessary to realize the service-oriented model resources. In combination with the characteristics of model resources, the service-oriented model resources can be completed through basic operations such as model encapsulation, model registration, model publishing, model discovery, model calling, and model orchestration. Among them, model encapsulation refers to encapsulating the model into a shared resource that is independent of the platform and can be flexibly deployed; model registration refers to registering the basic information of the model on the service platform and synchronously updating it to the model list for relevant users to consult; model publishing refers to publishing the model resources through the service platform, which supports the visibility of users across the network; model discovery refers to the ability of users across the network to search and discover model resources through the service platform; model calling refers to the ability of users to choose the most appropriate way to call the model according to the actual situation. It can be considered to download the model to the local computer and run it directly, or to use the model running on the service platform through the remote call interface to return the calculation results; model orchestration refers to the realization of task-oriented model application process, expressing the sequence, branch, loop and union or relationship between models, and adapting the model call data interface. Through the reasonable design of related operations, dynamic model orchestration can be realized to support both centralized services and edge adaptive services for distributed scenarios.
[0023] Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field and related fields without creative work should fall within the scope of protection of the present invention. The structures, devices and operating methods not specifically described and explained in the present invention are implemented according to the conventional means in the field unless otherwise specified and limited.
Claims
1. A task-driven command model dynamic orchestration system, characterized in that: It includes a resource construction module, which includes a command model, a command model knowledge graph unit and a task handling domain knowledge module; The command model is used to construct a combined generation model resource pool; The command model knowledge graph unit is used to realize the discovery and call of model resources; The task-driven command model dynamic orchestration system also includes: An organization and application module, which consists of a semantic analysis module, a task decomposition module, and a model arrangement module; The semantic parsing module is based on a large language model and accurately converts the user's natural language into unambiguous task instructions that can be understood by the computer through semantic parsing; The task decomposition module is used to decompose the user's instruction tasks layer by layer and match the decomposed instructions with the command model resources.
2. According to claim 1, a task-driven command model dynamic arrangement system is characterized by: The task handling domain knowledge module is used to match the corresponding model resources based on the analysis and calculation of the user description, and based on the workflow flexible modeling, hierarchical task network and other technologies, it describes the user's common business scenarios and corresponding command workflows, and establishes corresponding command activity decomposition rules.
3. The task-driven command model dynamic arrangement system according to claim 2, characterized in that: The model orchestration module is used to match model resources and analysis and computing tasks.