Method for designing military simulation agent based on knowledge graph

Abstract

The application relates to the technical field of artificial intelligence, and discloses a military simulation intelligent agent design method based on a knowledge graph, which comprises the following steps: taking a synchronous matrix, an execution matrix and decision support information formed in a military planning and war game process as the basis, uniformly representing units, stages, combat functions, decision points and condition elements in a knowledge graph, and forming a computable high-dimensional plan state expression; on the basis, a plan updating model reflecting stage promotion, function cooperation and decision point triggering relationship is constructed, which is used for describing dynamic evolution of a plan state in a simulation process; and through systematic depiction of a transient amplification phenomenon in a plan evolution process, the transient amplification phenomenon is clearly associated with decision points and condition elements in the knowledge graph, so that the simulation intelligent agent has stable and reproducible behavior performance in a decision generation and action execution process.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method for designing military simulation intelligent agents based on knowledge graphs. Background Technology

[0002] With the continuous development of information-based and system-of-systems warfare concepts, military simulation systems have evolved from simple model verification tools into a crucial technological foundation supporting operational planning, scheme evaluation, command training, and simulation analysis. In current military planning and simulation processes, wargaming and COA analysis are commonly used to repeatedly deduce and evaluate operational concepts under the commander's intent. During this process, structured results such as decision points, key events, condition triggers, and synchronization and execution matrices are widely used to describe the collaborative relationships of operational actions across time, space, and operational functions. This decision-point → condition → action-driven simulation approach has become an indispensable mechanism in modern military simulation systems, decisively influencing the validity and comparability of simulation results.

[0003] In recent years, knowledge graph technology has been increasingly introduced into the fields of military simulation and intelligent agent design to unify the representation of combat entities, mission relationships, phase divisions, command and control elements, and decision support information. Through knowledge graphs, multi-source structured information such as command documents, synchronization matrices, and decision support matrices can be integrated at the semantic level, providing a shared operational cognitive foundation for simulated intelligent agents. However, in simulation scenarios dominated by decision points and condition triggers, information updates in knowledge graphs often exhibit strong coupling, high dimensionality, and cross-phase propagation characteristics. When agents update their states and generate actions based on this information, minor changes in conditions, time deviations, or differences in the order of information injection can trigger significant changes in the planned state within a short period. This can lead to significant differences in the same combat plan across multiple simulation evaluations, affecting the stability of simulation analysis and engineering reliability.

[0004] From a computational and system dynamics perspective, military simulations triggered by decision points are essentially iterative updates of a high-dimensional planned state over discrete time. This update process involves multiple asymmetric coupling relationships, such as phase advancement, operational function coordination, and condition injection, which easily leads to non-normal linear operator structures. Under such structures, even if the system as a whole exhibits a stable state, significant transient amplification may still occur in a short period, making the planned state exceptionally sensitive to initial disturbances and changes in conditions. This phenomenon is particularly prominent in military simulation scenarios centered on synchronization matrices and decision support matrices. If it is not systematically characterized and constrained, it will weaken the practical value of knowledge graph-based military simulation agents in scheme evaluation, comparative analysis, and command decision support. Summary of the Invention

[0005] This invention provides a method for designing military simulation intelligent agents based on knowledge graphs, which solves the technical problems mentioned in the background art.

[0006] This invention provides a method for designing military simulation intelligent agents based on knowledge graphs, including:

[0007] Constructing a wargaming combat element map: Instantiating the collaborative action nodes and decision trigger condition nodes in the wargaming plan into map entities in the wargaming combat element map;

[0008] Constructing a combat situation evolution transmission model: Based on the topological connection relationship of the wargaming combat element map, a multi-dimensional battlefield situation state vector is generated, and a linear state transition matrix is ​​established to represent the combat phase advancement logic, combat function serial coordination logic, and decision trigger injection relationship. The linear state transition matrix constitutes the calculation kernel of the combat situation evolution transmission model.

[0009] Calculate the transient divergence index of the simulation results: Perform numerical stability calculation on the linear state transition matrix in the combat situation evolution transfer model, calculate the transient divergence index of the simulation results that characterizes the amplification of small disturbances, and extract the dominant divergence risk mode;

[0010] Generate a wargaming combat element map containing stability constraint attributes: map the dominant divergent risk mode into sensitivity attributes and triggering attributes, and write the sensitivity attributes and triggering attributes back to the corresponding map entities of the wargaming combat element map to obtain a wargaming combat element map containing stability constraint attributes.

[0011] Output military simulation agent: Construct a deterministic situational awareness sequence based on the numerical ranking of the sensitivity attribute and the triggering attribute in the wargaming combat element map containing stability constraint attributes, and generate agent action strategy by combining divergent suppression decision logic, thereby obtaining a military simulation agent with the ability to resist irregular transient amplification.

[0012] The beneficial effects of this invention include: mapping staff deduction products such as decision points, condition triggers, synchronization matrices, and execution matrices into a knowledge graph structure, and constructing a high-dimensional plan state update operator on this basis, systematically characterizing the transient amplification mechanism in the plan evolution process driven by decision points. This enables military simulation agents to maintain the reproducibility and engineering stability of simulation behavior and evaluation results when facing complex phase advancements and cross-functional collaboration. By introducing an information organization method directly related to plan update dynamics at the knowledge graph level, the distortion of results caused by nonlinear amplification of minor condition changes during simulation is avoided, thereby significantly improving the reliability and practical value of agent applications in combat plan analysis, wargaming, and command training. Attached Figure Description

[0013] Figure 1 This is a flowchart of the knowledge graph-based military simulation intelligent agent design method of the present invention. Detailed Implementation

[0014] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0015] like Figure 1 As shown, the knowledge graph-based military simulation intelligent agent design method includes:

[0016] Constructing a wargaming combat element map: Instantiating the collaborative action nodes and decision trigger condition nodes in the wargaming plan into map entities in the wargaming combat element map;

[0017] Constructing a combat situation evolution transmission model: Based on the topological connection relationship of the wargaming combat element map, a multi-dimensional battlefield situation state vector is generated, and a linear state transition matrix is ​​established to represent the combat phase advancement logic, combat function serial coordination logic, and decision trigger injection relationship. The linear state transition matrix constitutes the calculation kernel of the combat situation evolution transmission model.

[0018] Calculate the transient divergence index of the simulation results: Perform numerical stability calculation on the linear state transition matrix in the combat situation evolution transfer model, calculate the transient divergence index of the simulation results that characterizes the amplification of small disturbances, and extract the dominant divergence risk mode;

[0019] Generate a wargaming combat element map containing stability constraint attributes: map the dominant divergent risk mode into sensitivity attributes and triggering attributes, and write the sensitivity attributes and triggering attributes back to the corresponding map entities of the wargaming combat element map to obtain a wargaming combat element map containing stability constraint attributes.

[0020] Output military simulation agent: Construct a deterministic situational awareness sequence based on the numerical ranking of the sensitivity attribute and the triggering attribute in the wargaming combat element map containing stability constraint attributes, and generate agent action strategy by combining divergent suppression decision logic, thereby obtaining a military simulation agent with the ability to resist irregular transient amplification.

[0021] In one embodiment of the present invention, the collaborative action nodes and decision triggering condition nodes in the wargaming plan are instantiated as graph entities in the wargaming combat element graph, including:

[0022] Define the set of entities in a wargaming simulation: the set of combat units. Combat Phase Assembly Combat Function Set Decision point set and the set of triggering conditions ;

[0023] For any Generate a collaborative action node And establish the following graph relation edges:

[0024] ;

[0025] ;

[0026] ;

[0027] in, Indicates a relationship of ownership. Indicates the stage relationship. Indicates functional relationships;

[0028] For any Generate a decision trigger condition node And establish that the node points to the corresponding decision point. and triggering conditions The associated edges.

[0029] In one embodiment of the present invention, a multi-dimensional battlefield situation vector is generated based on the topological connectivity of the wargaming combat element map, including:

[0030] Using vectorized operators Mapping the state of the map entities in the wargaming combat element map to time points Multidimensional battlefield situational vector The calculation formula is as follows:

[0031]

[0032]

[0033] in: This is a cooperative action strength vector, where each dimension corresponds to a cooperative action node. At any moment The intensity of the execution state; This is a vector representing the degree to which trigger conditions are satisfied, with each dimension corresponding to a decision trigger condition node. At any moment The numerical value of condition satisfaction; The total dimension of the multidimensional battlefield situation vector; These represent the potential of the set of combat units, combat phases, combat functions, decision points, and triggering conditions, respectively.

[0034] In one embodiment of the present invention, a linear state transition matrix is ​​established to represent the operational phase progression logic, the serial coordination logic of combat functions, and the decision triggering injection relationship, including:

[0035] The following three basic component matrices are extracted from the wargaming combat element map:

[0036] Define the operational phase advance matrix Its elements satisfy:

[0037] ;

[0038] This indicates the logic of advancing one unit sequentially in the combat phase;

[0039] Define the combat function serial coordination matrix This matrix is ​​a fixed upper triangular matrix, representing a unidirectional serial dependency chain between combat functions as defined by wargaming rules;

[0040] Define the decision-triggered injection matrix If a path exists in the graph:

[0041] ;

[0042] And a path exists:

[0043] ;

[0044] Then place Otherwise, set it to 0.

[0045] In one embodiment of the present invention, the linear state transition matrix constitutes the computational kernel of the combat situation evolution transmission model, including:

[0046] Construct the linear state transition matrix The calculation formula is as follows:

[0047]

[0048] in: It is a combat unit identity matrix with a dimension equal to the total number of combat units; This is the advance matrix for the aforementioned operational phase; This is the serial coordination matrix for the aforementioned combat functions; Inject the matrix to trigger the decision; The Kronecker product operator is used to couple the logical structure of combat units, combat phases, and combat functions into a high-dimensional operator. A fixed scaling constant greater than zero is used to control the magnitude of the impact of decision triggers on the situation; the linear state transition matrix The non-zero structure characterizes the irregular coupling properties induced by decision points in wargaming.

[0049] In one embodiment of the present invention, the linear state transition matrix in the combat situation evolution transmission model is numerically stabilized, a transient divergence index characterizing the amplification degree of small disturbances in the simulation results is calculated, and the dominant divergence risk mode is extracted, including:

[0050] Define the linear state transition matrix. Pre-operator for:

[0051] ;

[0052] in For complex frequency variables, It is the identity matrix;

[0053] In the pre-defined complex plane grid area The search maximizes the dominant frequency of the weighted pre-solution operator norm. :

[0054] ;

[0055] in Representing complex numbers The real part, Represents the spectral norm of a matrix;

[0056] Calculate the transient divergence index of the inference results. :

[0057] ;

[0058] Pre-operator at the dominant frequency Perform singular value decomposition:

[0059] ;

[0060] Extract the left singular vector corresponding to the maximum singular value. With right singular vectors ,Will and The combination is defined as the dominant divergent risk mode.

[0061] In one embodiment of the present invention, the dominant divergent risk mode is mapped to a sensitivity attribute and a triggering attribute, and the sensitivity attribute and the triggering attribute are written back to the corresponding map entity of the wargaming combat element map to obtain a wargaming combat element map containing stability constraint attributes, including:

[0062] Constructing a cooperative action node selection matrix Decision trigger condition node selection matrix Calculate the sensitivity attribute vector With triggering attribute vector :

[0063] ;

[0064] ;

[0065] in: This indicates that the absolute value operation is performed on each element of the vector. The output mode vector, The input mode vector;

[0066] Traverse the wargaming simulation combat element map and convert the vectors Each element's value is assigned to the corresponding collaborative action node. This is denoted as the sensitivity attribute of the entity;

[0067] vector Each element's value is assigned to the corresponding decision triggering condition node. This is denoted as the triggering attribute of the entity;

[0068] This generates a wargaming combat element map that includes stability constraint attributes.

[0069] In one embodiment of the present invention, a deterministic situational awareness sequence is constructed based on the numerical ranking of the sensitivity attribute and the triggering attribute in the wargaming combat element map containing stability constraint attributes, including:

[0070] Construct the feather sorting permutation matrix And use this matrix to represent the multidimensional battlefield situation state vector. Transformed into the deterministic situational awareness sequence The calculation formula is as follows:

[0071] ;

[0072] in: Based on trigger attribute vector The decision trigger condition node index sequence obtained by arranging the element values ​​from largest to smallest; Based on sensitivity attribute vector The sequence of indexes of collaborative action nodes obtained by arranging the element values ​​from largest to smallest; This indicates an operation that rearranges the vector elements according to their index sequence; The degree vector is the condition for triggering the event. This is the intensity vector of the coordinated action.

[0073] In one embodiment of the present invention, a military simulation agent with resistance to irregular transient amplification is obtained by combining divergent suppression decision logic to generate an agent action strategy, including:

[0074] Define the action selection policy function for the agent and compute each candidate action. Action confidence :

[0075] ;

[0076] in: Candidate actions Embedded vectors of map entities in the wargaming combat element map; For the input mapping parameter matrix, This refers to the deterministic situational awareness sequence; The divergence suppression weight coefficient is greater than zero;

[0077] The divergence risk cost of this action is defined as follows:

[0078] ;

[0079] in: and Actions The set of collaborative action nodes and the set of decision triggering condition nodes associated in the graph;

[0080] The agent outputs actions Using deterministic maximization selection:

[0081] .

[0082] Combat units refer to all red and blue forces and neutral combat entities participating in wargaming. They are the basic organizational units that constitute wargaming, such as the first mechanized infantry brigade, fire support battalion, reconnaissance company, etc. They need to be clearly defined and included in the combat unit set before the wargaming.

[0083] Operational phases refer to operational links divided according to time sequence. They are segments of operational actions in the time dimension, such as the assembly and deployment phase, the obstacle breaching and assault phase, and the in-depth expansion phase. The sequence of phases needs to be predetermined in the simulation plan.

[0084] Combat functions refer to the classification of business areas according to operational doctrines. They are the functional divisions of combat operations, such as intelligence, mobility, firepower, support, command and control, etc. The function types must comply with the rules and requirements of wargaming.

[0085] Decision points refer to the critical nodes in the course of combat that require commanders to make judgments, which directly affect the adjustment and advancement of combat plans. Examples include the timing of enemy reserve deployment and the timing of breakthroughs in forward positions. Key decision points need to be identified in conjunction with simulation scenarios.

[0086] Triggering conditions refer to the prerequisites for activating a decision point and are the basis for decision point judgment. Examples include discovering an enemy tank battalion or when one's own ammunition is more than half depleted. Each decision point must correspond to at least one triggering condition.

[0087] A collaborative action node is a logical container that represents a specific combination of combat unit, combat phase, and combat function. It is generated by traversing all the triplets of combat unit, combat phase, and combat function. Each combination uniquely corresponds to a collaborative action node. For example, the collaborative action node of the First Mechanized Infantry Brigade performing intelligence function during the assembly and deployment phase.

[0088] Graph association edges (attribution relationships) refer to graph edges that connect collaborative action nodes with their corresponding combat units. The type of edge is defined as an attribution relationship, and its function is to clarify the combat unit to which each collaborative action node belongs in the organizational structure.

[0089] Graph association edges (phase associations) refer to graph edges that connect collaborative action nodes to their corresponding combat phases. The type of edge is defined as a phase association, and its function is to clarify the combat phase corresponding to each collaborative action node in the time dimension.

[0090] Graph association edges (functional associations) refer to graph edges that connect collaborative action nodes with their corresponding combat functions. The type of edge is defined as a functional association, and its function is to clarify the combat function corresponding to each collaborative action node in the functional dimension.

[0091] A decision trigger condition node is a logical container that represents a specific decision point-trigger condition combination. It is generated by traversing all valid binary combinations of decision points and trigger conditions (only retaining combinations that actually exist in the simulation plan). Each valid combination uniquely corresponds to a decision trigger condition node, such as the decision trigger condition node for the combination of the enemy reserve deployment time and the discovery of the enemy tank battalion.

[0092] Decision trigger condition node association edges refer to graph edges that connect decision trigger condition nodes with their corresponding decision points and corresponding trigger conditions. They do not have a separate edge type name, and their purpose is to clarify the decision point and trigger condition corresponding to each decision trigger condition node and establish the logical relationship between the two.

[0093] A fixed order of graph node unfolding refers to defining a fixed rule for arranging graph nodes in a vector. This rule needs to be determined at the beginning of the simulation and remains unchanged throughout the simulation. For example, coordinate action nodes can be sorted according to the lexicographical order of combat unit, combat phase, and combat function, and decision trigger condition nodes can be sorted according to the lexicographical order of decision point and trigger condition, ensuring that the correspondence between vector dimension and node is always consistent.

[0094] The collaborative action intensity vector is a vector that represents the execution state intensity of all collaborative action nodes. Each element in the vector corresponds to the execution state intensity of a collaborative action node at a certain moment. The intensity value is a continuous real number from 0 to 1, where 0 indicates that the node has not started execution, 1 indicates that the node has fully executed, and 0.5 indicates that the node has executed more than half of its tasks. The value needs to be updated in real time according to the simulation process.

[0095] The trigger condition satisfaction vector is a vector that represents the degree of satisfaction of all decision trigger condition nodes. Each element in the vector corresponds to the satisfaction value of a decision trigger condition node at a certain moment. The value can be represented by a Boolean floating-point value of 0 or 1 (0 for not satisfied, 1 for satisfied), or it can be a confidence probability value between 0 and 1 based on fuzzy logic calculation. The value needs to be updated in real time according to the inference process.

[0096] The multidimensional battlefield situation vector refers to a high-dimensional vector that integrates the coordinated action intensity vector and the trigger condition satisfaction vector. It is generated by vertically concatenating the coordinated action intensity vector and the trigger condition satisfaction vector. That is, first, all elements of the coordinated action intensity vector are arranged completely, and then all elements of the trigger condition satisfaction vector are arranged completely, forming a high-dimensional vector that covers the state of all nodes.

[0097] Vectorization operators are computational tools that map the state of graph entities into vectors. The specific computational logic is to extract the execution intensity value of each cooperative action node and the satisfaction value of each decision trigger condition node in a fixed order of graph node expansion, and then arrange them in order to form a multi-dimensional battlefield situation state vector.

[0098] A moment refers to the time node corresponding to the situation vector, that is, the time step in wargaming, such as the first time step, the second time step, etc. Each time step corresponds to a multi-dimensional battlefield situation vector, which is used to record the complete battlefield situation at that time point.

[0099] The total dimension refers to the total number of dimensions of the multidimensional battlefield situation vector. It is calculated as the sum of the total number of coordinated action nodes and the total number of decision trigger condition nodes. The total number of coordinated action nodes is equal to the product of the number of elements in the set of combat units, the set of combat phases, and the set of combat functions. The total number of decision trigger condition nodes is equal to the product of the number of elements in the set of decision points and the set of trigger conditions.

[0100] The cardinality of a set refers to the number of elements in each entity set, that is, the number of specific entities contained in each set. For example, if the combat unit set has 3 entities (1st Mechanized Infantry Brigade, Fire Support Battalion, and Reconnaissance Company), then the cardinality of the set is 3. The number of elements in each set needs to be counted and the cardinality of the set needs to be determined before the simulation.

[0101] The total number of coordinated action nodes refers to the total number of all coordinated action nodes. It is calculated by multiplying the number of elements in the combat unit set, the number of elements in the combat phase set, and the number of elements in the combat function set. For example, if there are 3 combat units, 3 combat phases, and 5 combat functions, then the total number of coordinated action nodes is 3×3×5=45.

[0102] The total number of decision trigger condition nodes refers to the total number of decision trigger condition nodes. It is calculated by multiplying the number of elements in the decision point set by the number of elements in the trigger condition set. For example, if there are 2 decision points and 2 trigger conditions, then the total number of decision trigger condition nodes is 2×2=4.

[0103] The operational phase advancement matrix is ​​a matrix that represents the logical evolution of the operational phases. The number of rows and columns of the matrix are equal to the number of elements in the operational phase set. The matrix elements can only be 0 or 1. When the column index is equal to the row index plus 1, the element value is 1 (indicating that the current phase can advance to the next phase). Otherwise, the element value is 0. For example, when there are 3 operational phases, the elements in the first row and second column and the second row and third column of the matrix are 1, and the rest are 0.

[0104] The combat function serial coordination matrix is ​​a matrix that represents the unidirectional dependency relationship between combat functions. The number of rows and columns of the matrix are equal to the number of elements in the combat function set. The matrix is ​​a fixed upper triangular matrix (excluding the diagonal). The position of the non-zero element corresponds to the unidirectional dependency chain defined by the wargaming rules. For example, the intelligence function corresponds to index 1 and the maneuver function corresponds to index 2. If the rule requires intelligence to be executed before maneuver, then the element in the first row and second column of the matrix is ​​1, which is used to clarify the sequential execution logic between functions.

[0105] The decision trigger injection matrix is ​​a matrix that represents the influence of decision trigger conditions on collaborative actions. The number of rows in the matrix is ​​equal to the total number of collaborative action nodes, and the number of columns is equal to the total number of decision trigger condition nodes. The matrix elements can only be 0 or 1. When a decision trigger condition node and a collaborative action node have an effective influence path, the corresponding element is 1, otherwise it is 0.

[0106] The phase sequence refers to the time order of the combat phases in the war game simulation plan, such as assembly and deployment → obstacle breaching and assault → in-depth expansion. It is the core basis for constructing the combat phase advancement matrix, and the order of the phases needs to be clarified and solidified before the simulation.

[0107] The order of combat function dependence refers to the sequential execution logic of different combat functions within the same operational phase. For example, within the same phase, intelligence functions are executed first, followed by maneuver functions, and finally firepower functions. This is the core basis for constructing a sequential coordination matrix of combat functions and must conform to the tactical rules of wargaming.

[0108] The graph path search results refer to the search results of the node association paths in the wargaming combat element graph. By traversing the nodes and edges in the graph, all possible association paths from the decision trigger condition node to the cooperative action node are found. This is the basis for determining whether there are effective influence paths.

[0109] An effective influence path refers to a valid association path from a decision trigger condition node to a collaborative action node. It must satisfy two conditions simultaneously: first, the decision point points to an action entity through a trigger relationship, and the action entity points to a collaborative action node through an influence relationship; second, the action entity points to the trigger condition through a dependency relationship. Only when both sub-paths are met is it considered an effective influence path.

[0110] The effective influence value refers to the numerical value in the decision trigger injection matrix that represents the existence of an effective influence. The value is fixed at 1. When there is an effective influence path between the decision trigger condition node and the collaborative action node, the corresponding position in the matrix is ​​marked as 1, and when there is no effective influence path, it is marked as 0.

[0111] A path element refers to a decision point that effectively influences a path, i.e., the decision point entity at the beginning of the path. It is a specific element in the set of decision points, such as the decision point at the moment when the enemy reserves are deployed, which is the starting point of the effective influence path.

[0112] Path elements refer to action entities that effectively influence a path. They are intermediate nodes that connect decision points with collaborative action nodes and triggering conditions. Examples include specific actions such as conducting battlefield reconnaissance and launching fire strikes. In the graph, they exist as independent entities and have relationships such as triggering, influencing, and depending.

[0113] A path element refers to a triggering condition that effectively affects a path, that is, a triggering condition entity on which an action entity depends. It is a specific element in the set of triggering conditions. For example, the triggering condition of discovering an enemy tank battalion is a prerequisite for the action entity to execute.

[0114] The linear state transition matrix is ​​the computational kernel of the operational situation evolution transmission model. It is a high-dimensional matrix that integrates the natural evolution of the situation and the influence of decision intervention. It is formed by superimposing the dominant driving term and the triggering injection term. Its function is to depict the dynamic evolution logic of the battlefield situation in wargaming, including the changes caused by natural advancement and decision triggers.

[0115] The combat unit identity matrix is ​​a square matrix with dimensions equal to the total number of combat units. The matrix type is an identity matrix, meaning that all elements on the diagonal are 1 and all elements off the diagonal are 0. Its function is to represent the identity of combat units and ensure that combat units do not switch or become confused without reason during the evolution of the situation.

[0116] Kronecker product operation refers to a mathematical operation that couples low-dimensional matrices into high-dimensional operators. Specifically, it is used to couple the combat unit identity matrix, the combat phase advancement matrix, and the combat function serial coordination matrix through tensor product to form a high-dimensional matrix that covers the state transition logic of all coordinated actions, ensuring that the three-dimensional low-dimensional logic (unit, phase, function) is fully integrated into the high-dimensional evolution model.

[0117] The dominant advancement term refers to the matrix term that characterizes the natural evolution of the multidimensional battlefield situation under undisturbed conditions. It is generated by coupling the combat unit identity matrix, the combat phase advancement matrix, and the combat function serial coordination matrix through Kronecker product operation. Its function is to describe the natural development of the battlefield situation along the phase sequence and function dependence when there is no decision intervention.

[0118] Block matrix concatenation refers to the operation of combining matrix blocks with different functions into a complete matrix according to a specific structure. In this method, it is used to superimpose bias blocks onto a specified region of the dominant propulsion term to ensure that the impact of decision triggering is accurately incorporated into the linear state transition matrix.

[0119] The fixed scaling constant is a preset positive parameter that controls the magnitude of the impact of the decision trigger on the situation. The preferred value range is 1.0 to 10.0. The specific value needs to be set according to the time step sensitivity of the wargaming system to ensure that the decision trigger signal is numerically sufficient to cause a significant change in the situation, while avoiding being overwhelmed by calculation errors.

[0120] The bias block is a matrix block obtained by multiplying the decision trigger injection matrix by a fixed scaling constant. The matrix structure is a two-row, two-column block matrix, in which the upper left, lower left, and lower right blocks are all zero matrices, and only the upper right block is the decision trigger injection matrix multiplied by the fixed scaling constant. Its function is to inject the impact of decision triggers into the situation evolution model.

[0121] The upper right block region of the dominant driving term refers to the block region located in the upper right position in the dominant driving term matrix. It is the only superposition position of the bias block. The core basis for selecting this position is to ensure that the impact path of decision triggering on coordinated actions is clear and conforms to the causal logic of decision-condition-action.

[0122] The upper triangular structure refers to the matrix structure characteristics of a linear state transition matrix. That is, after the matrix is ​​divided into blocks, only the upper triangular region blocks may have non-zero elements, while the lower triangular region blocks are all zero matrices. This structure is used to ensure the temporal irreversibility and logical order of situation evolution.

[0123] The superdiagonal coupling feature refers to the fact that in the linear state transition matrix, non-zero elements not only exist in the main diagonal block, but also in the cross-block region above the main diagonal. This feature is used to characterize the cross-stage and cross-functional situational coupling relationship triggered by decision-making.

[0124] Discontinuous situational transitions refer to the phenomenon of sudden changes in battlefield situation triggered by decision points. That is, the situation does not evolve gradually along the natural path, but rather undergoes a leapfrog change due to decision intervention. The linear state transition matrix accurately describes this phenomenon through the superposition of bias blocks.

[0125] Transient evolution refers to the rapid changes in the battlefield situation within a short period of time, especially the short-term process of the situation transitioning from one stable state to another after a decision is triggered. The irregular structure of the linear state transition matrix can accurately capture the dynamic characteristics of this process.

[0126] The complex frequency domain refers to the frequency space used for numerical stability calculations. This space includes a real part and an imaginary part. The real part represents the growth or decay trend of the situation change, while the imaginary part represents the oscillation frequency of the situation change. Its function is to provide an analytical scenario for calculating the degree of transient amplification.

[0127] A fixed discrete scanning grid refers to a discrete sampling grid constructed in the complex frequency domain, used to comprehensively search for frequency points that may cause the maximum transient amplification. The real part of the grid is preferably set to 0 to 2.0 with a step size of 0.1 (covering the transient region from edge stability to rapid growth), and the imaginary part is preferably set to -π to π with a step size of 0.1 (covering the oscillation frequency range that may occur in the wargaming system). Once the grid structure is determined, it remains unchanged throughout the entire process.

[0128] Complex frequency points refer to a single sampling point in a fixed discrete scanning grid. Each complex frequency point consists of a real part and an imaginary part. It is the basic unit for calculating transient amplification factor. Traversing all complex frequency points can cover all possible situational evolution frequency scenarios.

[0129] A pre-operator is a mathematical operator used to calculate the transient amplification factor at a specific complex frequency. It is obtained by multiplying the identity matrix by the current complex frequency and then subtracting the linear state transition matrix to get a new matrix. The inverse operation of this new matrix is ​​the pre-operator, which is used to quantify the system's response to disturbances at that frequency.

[0130] The spectral norm is a matrix norm used to measure the maximum singular value of a matrix. It is a quantitative indicator of the maximum amplification capability of a pre-determiner at a specific complex frequency point for a unit perturbation. It is calculated by performing singular value decomposition on the pre-determiner and taking the maximum singular value obtained after decomposition, which is the spectral norm.

[0131] Transient amplification factor refers to the degree to which a system amplifies small disturbances at a specific complex frequency point. Numerically, it is equal to the spectral norm of the pre-solution operator at that frequency point. The larger the amplification factor, the more significant the impact of the disturbance on the situation at that frequency point.

[0132] A specific complex frequency point refers to the complex frequency point in a fixed discrete scanning grid that maximizes the product of the transient amplification factor and the real part of that frequency point. This frequency point is the most dangerous frequency that determines the transient amplification behavior of the system. It is obtained by traversing all complex frequency points in the discrete scanning grid, calculating the product one by one, and filtering out the frequency point corresponding to the maximum value.

[0133] Logarithmic compression is a mathematical operation that converts exponentially increasing transient magnification into linear, readable values. It uses natural logarithm operations to avoid numerical overflow or incomparability caused by excessive magnification, while preserving the relative differences in magnification.

[0134] The transient divergence index of the simulation results refers to the core indicator of the amplification degree of small disturbances in the quantitative simulation results. It is calculated by multiplying the real part of a specific complex frequency point by the spectral norm of the pre-solution operator at that frequency point, adding 1 (to prevent meaningless smoothing terms in logarithmic operations), and then performing logarithmic compression on the result. The larger the index value, the higher the sensitivity of the simulation results to the initial disturbances and the greater the divergence risk.

[0135] The identity matrix is ​​an identity matrix with the same dimensions as the linear state transition matrix. Its structure consists of diagonal elements all being 1 and off-diagonal elements all being 0. Its function is to participate in the construction of pre-solution operators to ensure dimensional matching of matrix operations.

[0136] The real part of a complex number refers to the part of the complex number frequency that does not include the imaginary unit. It represents the growth or decay trend of the situation change at that frequency point. The larger the value of the real part, the more likely the situation is to show growth at that frequency point.

[0137] Singular value decomposition (SVD) is a mathematical operation that decomposes a pre-determined operator at a specific complex frequency point into three matrices: a left orthogonal matrix, a singular value diagonal matrix, and a right orthogonal matrix. Its purpose is to extract the maximum singular value of the pre-determined operator and its corresponding vector, thereby locating the sensitive direction and triggering direction of the system.

[0138] The maximum singular value refers to the largest value at the diagonal position in the singular value diagonal matrix obtained after singular value decomposition. This value is equal to the spectral norm of the pre-decomposer and is a direct reflection of the system's maximum amplification capability at that frequency.

[0139] The left singular vector refers to the column vector corresponding to the largest singular value in the left orthogonal matrix obtained after singular value decomposition, that is, the first column of the left orthogonal matrix. Its function is to characterize the sensitive direction of the system output, that is, the direction of the situation state that is most easily amplified by disturbance.

[0140] The right singular vector is the column vector corresponding to the largest singular value in the right orthogonal matrix obtained after singular value decomposition, that is, the first column of the right orthogonal matrix. Its function is to characterize the triggering direction of the system input, that is, the perturbation input direction that is most likely to cause global divergence.

[0141] The output mode vector refers to the vector that represents the sensitive direction of the system output in the dominant divergent risk mode. It is completely consistent with the left singular vector and is a functional definition of the left singular vector, used to clarify its role in divergent risk analysis.

[0142] The input mode vector refers to the vector that represents the direction of system input triggering in the dominant divergent risk mode. It is completely consistent with the right singular vector and is a functional definition of the right singular vector, used to clarify its role in divergent risk analysis.

[0143] The dominant divergence risk mode refers to the risk representation carrier formed by the combination of the output mode vector (left singular vector) and the input mode vector (right singular vector). Its function is to fully characterize the most dangerous divergence risk of the system. The input mode vector clarifies which disturbances will cause the maximum divergence, and the output mode vector clarifies which situations will be amplified to the greatest extent.

[0144] The cooperative action node selection matrix is ​​a sparse matrix that accurately extracts the relevant components of cooperative action nodes from the full state vector. The number of rows in the matrix is ​​equal to the total number of cooperative action nodes, and the number of columns is equal to the total dimension of the multi-dimensional battlefield situation state vector. The matrix structure is that the column positions corresponding to the cooperative action nodes form an identity matrix, and the other column positions are all 0. Its function is to separate the state components related to cooperative actions from the high-dimensional state vector.

[0145] The decision trigger condition node selection matrix is ​​a sparse matrix that accurately extracts the relevant components of decision trigger condition nodes from the full state vector. The number of rows in the matrix is ​​equal to the total number of decision trigger condition nodes, and the number of columns is equal to the total dimension of the multi-dimensional battlefield situation state vector. The matrix structure is that the column positions corresponding to the decision trigger condition nodes form an identity matrix, and the other column positions are all 0. Its function is to separate the state components related to decision trigger conditions from the high-dimensional state vector.

[0146] The sensitivity attribute vector is a vector that represents the divergence risk level of all collaborative action nodes. It is obtained by performing matrix multiplication between the collaborative action node selection matrix and the output mode vector (left singular vector), and taking the absolute value of each element of the result. The resulting vector is the sensitivity attribute vector, and each element in the vector corresponds to the basic value of the divergence risk of a collaborative action node.

[0147] The trigger attribute vector is a vector that represents the divergence risk level of all decision trigger condition nodes. It is obtained by performing matrix multiplication between the decision trigger condition node selection matrix and the input modality vector (right singular vector), and taking the absolute value of each element of the result. The resulting vector is the trigger attribute vector, and each element in the vector corresponds to the basic value of the divergence risk of a decision trigger condition node.

[0148] The element-wise absolute value operation is a mathematical operation that eliminates the influence of the sign of each element in a vector. The operation rule is to retain the numerical value of the element and ignore the positive or negative sign. Its purpose is to eliminate the interference of the oscillation sign in the vector on risk assessment and retain only the amplitude information of the risk level.

[0149] Sensitivity attribute refers to the node attribute that characterizes the risk level of a cooperative action node being erroneously amplified in the simulation. The value is taken from the element value of the corresponding cooperative action node in the sensitivity attribute vector. After being obtained, it is directly written into the corresponding cooperative action node in the war game simulation combat element map. The larger the value, the more easily the cooperative action is erroneously amplified by small disturbances.

[0150] Triggering attributes refer to node attributes that characterize the potential intensity of the divergence of the overall simulation results induced by decision triggering condition nodes. The values ​​are taken from the element values ​​of the corresponding decision triggering condition nodes in the triggering attribute vector. After being obtained, they are directly written into the corresponding decision triggering condition nodes in the wargaming simulation combat element map. The larger the value, the more likely the decision triggering condition is to cause the global simulation results to diverge.

[0151] The collaborative action node index refers to the unique position identifier of the collaborative action node in the multi-dimensional battlefield situation vector. It is obtained by assigning a unique number to each collaborative action node when defining a fixed map node unfolding order in the early stage of the simulation. Its function is to establish a one-to-one correspondence between sensitivity attribute vector elements and collaborative action nodes, ensuring that attribute values ​​are accurately written back.

[0152] The decision trigger condition node index refers to the unique position identifier of the decision trigger condition node in the multi-dimensional battlefield situation state vector. It is obtained by assigning a unique serial number to each decision trigger condition node when defining a fixed map node unfolding order in the early stage of the simulation. Its function is to establish a one-to-one correspondence between trigger attribute vector elements and decision trigger condition nodes, ensuring that attribute values ​​are accurately written back.

[0153] A wargaming combat element map containing stability constraint attributes refers to an enhanced map obtained by writing back sensitivity attributes and triggering attributes to the original wargaming combat element map. It is a combination of the original map and stability risk information. Its function is to provide risk constraint basis for the subsequent situational awareness and decision generation of the agent, enabling the agent to identify high-risk nodes.

[0154] A deterministic situational awareness sequence is a battlefield situational input sequence ordered by divergent risk priority. It is the actual input layer vector of the agent's neural network and is generated by splicing the trigger condition state subsequence and the cooperative action state subsequence in a fixed order. Its function is to ensure that the agent prioritizes the perception and processing of high-risk situational information that has the greatest impact on global stability, and to eliminate topological randomness at the input level.

[0155] The feather sorting permutation matrix is ​​a square matrix that sorts the multidimensional battlefield situational state vector according to divergent risk. Its dimension is consistent with the total dimension of the multidimensional battlefield situational state vector. Its function is to rearrange the elements in the original state vector according to risk level through matrix operations to generate a deterministic situational awareness sequence. In engineering implementation, the complete matrix does not need to be stored, and the rearrangement operation can be performed directly through the index sequence.

[0156] The decision trigger condition node index sequence refers to the sequence of decision trigger condition node position identifiers obtained by arranging the trigger attribute values ​​from largest to smallest. The method of obtaining it is to extract the trigger attribute values ​​of all decision trigger condition nodes in the graph containing stability constraint attributes, sort them in descending order of value, and record the index of the corresponding node. Its purpose is to clarify the element order of the trigger condition state subsequence.

[0157] The cooperative action node index sequence refers to the sequence of cooperative action node position identifiers obtained by arranging the sensitivity attribute values ​​from largest to smallest. It is obtained by extracting the sensitivity attribute values ​​of all cooperative action nodes in the graph containing stability constraint attributes, arranging them in descending order of value, and recording the index of the corresponding node. Its purpose is to clarify the element order of the cooperative action state subsequence.

[0158] The vector element rearrangement operation refers to the operation of adjusting the order of vector elements according to the index sequence. Specifically, for the trigger condition satisfaction vector, the element positions are rearranged according to the decision trigger condition node index sequence; for the cooperative action strength vector, the element positions are rearranged according to the cooperative action node index sequence to ensure that the vector element corresponding to the high-risk node is at the beginning of the sequence.

[0159] The trigger condition state subsequence refers to the trigger condition satisfaction vector rearranged according to the decision trigger condition node index sequence. It is generated by performing a vector element rearrangement operation on the original trigger condition satisfaction vector. Its purpose is to present the trigger condition states with high trigger risk in a concentrated manner, so that the agent can perceive them first.

[0160] The cooperative action state subsequence refers to the cooperative action intensity vector rearranged according to the cooperative action node index sequence. It is generated by performing a vector element rearrangement operation on the original cooperative action intensity vector. Its purpose is to present the cooperative action states with high sensitivity and risk in a concentrated manner, so that the agent can focus on them.

[0161] The fixed order refers to the fixed rule for assembling the trigger conditional state subsequence and the cooperative action state subsequence. Specifically, the trigger conditional state subsequence is executed first, followed by the cooperative action state subsequence. This order cannot be changed throughout the process. Its purpose is to ensure that the agent always prioritizes processing high-risk information at the decision-making trigger level before processing high-risk information at the action execution level.

[0162] Battlefield situation information with high divergence risk refers to the situation data at the front end of the deterministic situation awareness sequence. Specifically, it includes the satisfaction information of high-trigger attribute nodes in the first half of the trigger condition state subsequence and the execution intensity information of high-sensitivity attribute nodes in the first half of the cooperative action state subsequence. Its role is to provide the agent with the most critical risk situation input and support anti-divergence decision-making.

[0163] Graph embedding technology refers to a pre-training technique that maps action entities in a wargaming simulation's operational element graph into fixed-dimensional real vectors. Specifically, it can employ algorithms such as TransE, Node2Vec, or graph convolutional networks. Its function is to transform discrete action entities in the graph into computer-computable vector representations, providing a foundation for calculating action matching degrees.

[0164] Action entities refer to specific tactical actions that agents can perform in wargaming, such as launching fire strikes, conducting battlefield reconnaissance, and adjusting deployment positions. They exist as independent entities in the graph and have association relationships with cooperative action nodes and decision trigger condition nodes. They are the core components of the agent's action strategy.

[0165] An embedding vector is a fixed-dimensional real vector obtained by transforming an action entity through graph embedding technology. The vector dimension can be set according to the graph size (preferably 128 to 512 dimensions). Each action entity uniquely corresponds to an embedding vector, which is used to quantify the semantic features of the action entity and support the calculation of the matching degree with the latent state of the situation.

[0166] Candidate actions refer to the set of all action entities that an agent can choose to execute at the current moment. The range of the set is determined by the wargaming rules and the current battlefield situation. For example, when the agent is in the assembly phase, candidate actions may include troop assembly, material resupply, etc. The purpose is to provide the agent with a range of alternative actions to choose from.

[0167] Action confidence is a score that measures the priority of candidate actions. It is a comprehensive indicator that integrates situational matching and divergence risk penalty. Its purpose is to provide a quantitative basis for the agent's action selection. The higher the value, the more efficient and robust the action is.

[0168] The input mapping parameter matrix refers to the learnable parameter matrix that maps the deterministic situational awareness sequence to the action embedding space. Its dimension is the embedding vector dimension × the deterministic situational awareness sequence dimension. Its function is to establish the association between situational input and action features. The parameters are optimized and updated through the backpropagation algorithm during the agent training process.

[0169] The divergence suppression weight coefficient is a preset positive parameter that adjusts the penalty for divergence risk. The preferred value range is 0.5 to 5.0. The specific value is set according to the stability tolerance of the simulation system (a larger value is used for rigorous scheme evaluation scenarios and a smaller value is used for exploratory simulation scenarios). Its function is to balance the action benefits and divergence risks.

[0170] The divergence risk cost refers to the total divergence risk value of the nodes associated with the candidate action. It is calculated by summing the sensitivity attribute values ​​of all collaborative action nodes associated with the action, as well as the triggering attribute values ​​of all associated decision triggering condition nodes. Its purpose is to quantify the potential risk of global divergence caused by the candidate action.

[0171] The set of associated collaborative action nodes refers to the set of collaborative action nodes that are directly associated with the candidate action in the graph through influence relationships. It is obtained by searching the graph for valid paths from the action entity to the collaborative action node. Its purpose is to clarify the range of high-sensitivity risk nodes that the action may affect.

[0172] The set of associated decision trigger condition nodes refers to the set of decision trigger condition nodes that are directly associated with candidate actions in the graph through triggering or dependency relationships. It is obtained by searching the graph for valid paths from the action entity to the decision trigger condition nodes. Its purpose is to clarify the range of high-risk trigger nodes that the action may involve.

[0173] The latent state refers to the intermediate state vector obtained after the deterministic situational awareness sequence is transformed by the input mapping parameter matrix. It is a condensed representation of situational information and its function is to simplify the calculation of the matching degree between the situational and action embedding vectors and highlight key situational features.

[0174] Matching degree refers to the semantic fit between the hidden state and the action embedding vector. It is calculated by the inner product operation of the hidden state vector and the action embedding vector. The larger the value, the stronger the adaptability of the action to the current battlefield situation. Its function is to quantify the tactical effectiveness of the action.

[0175] The weighted divergence risk cost refers to the product of the divergence risk cost and the divergence suppression weight coefficient. It is calculated by multiplying the divergence risk cost of the candidate action by the preset divergence suppression weight coefficient. Its function is to convert the risk cost into a deduction item of the action confidence according to the set strength, thereby achieving risk suppression.

[0176] Final confidence score refers to the final priority score of candidate actions. It is calculated by subtracting the weighted divergence risk cost from the matching degree. Its purpose is to comprehensively consider the tactical effectiveness and stability of actions, and to provide the agent with a unique basis for action selection.

[0177] An agent's action strategy refers to the decision-making output plan of an agent in the current situation. Specifically, it involves selecting the action with the highest confidence from the set of candidate actions. Its purpose is to ensure that the agent executes actions that are highly adaptable and low-risk, and to have the ability to resist irregular transient amplification.

[0178] Output action refers to the unique action that the agent will execute in the next moment, that is, the specific action entity corresponding to the agent's action strategy. Its function is to transform the decision result into an executable tactical action and drive the evolution of the wargaming situation.

[0179] Deterministic maximization selection refers to a strategy that selects the action with the highest final confidence from the set of candidate actions. The specific logic is to iterate through the final confidence of all candidate actions and select the action with the highest value as the output action. Its purpose is to eliminate the randomness of action selection, ensure consistent output results under the same input conditions, and improve the reproducibility of simulation.

[0180] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A method for designing military simulation intelligent agents based on knowledge graphs, characterized in that, include: Constructing a wargaming combat element map: Instantiating the collaborative action nodes and decision trigger condition nodes in the wargaming plan into map entities in the wargaming combat element map; Constructing a combat situation evolution transmission model: Based on the topological connection relationship of the wargaming combat element map, a multi-dimensional battlefield situation state vector is generated, and a linear state transition matrix is ​​established to represent the combat phase advancement logic, combat function serial coordination logic, and decision trigger injection relationship. The linear state transition matrix constitutes the calculation kernel of the combat situation evolution transmission model. Calculate the transient divergence index of the simulation results: Perform numerical stability calculation on the linear state transition matrix in the combat situation evolution transfer model, calculate the transient divergence index of the simulation results that characterizes the amplification of small disturbances, and extract the dominant divergence risk mode, including: A fixed discrete scanning grid is constructed in the complex frequency domain. For each complex frequency point in the grid, the pre-operator of the linear state transition matrix is ​​calculated, and the spectral norm of the pre-operator is solved to obtain the transient amplification factor at that complex frequency point. The discrete scanning grid is traversed to find a specific complex frequency point where the product of the transient amplification factor and the real part of that frequency point reaches its maximum value. This maximum value is logarithmically compressed and used as the transient divergence index of the derivation result. Singular value decomposition is performed on the pre-operator at the specific complex frequency point. The left singular vector corresponding to the maximum singular value is extracted as the output mode vector in the dominant divergence risk mode representing the system output sensitive direction, and the right singular vector corresponding to the maximum singular value is extracted as the input mode vector in the dominant divergence risk mode representing the system input trigger direction. Generate a wargaming combat element map that includes stability constraint attributes: The dominant divergent risk mode is mapped to sensitivity attributes and triggering attributes, including: From the output mode vector of the dominant divergent risk mode, extract the vector elements corresponding to the cooperative action node index, and take the corresponding absolute value as the sensitivity attribute; From the input mode vector of the dominant divergent risk mode, extract the vector elements corresponding to the decision triggering condition node index, and take the corresponding absolute value as the triggering attribute; The sensitivity attribute and the triggering attribute are then written back to the corresponding map entity of the wargaming combat element map to obtain a wargaming combat element map containing stability constraint attributes. Output military simulation agent: Construct a deterministic situational awareness sequence based on the numerical ranking of the sensitivity attribute and the triggering attribute in the wargaming combat element map containing stability constraint attributes, and generate agent action strategy by combining divergent suppression decision logic, thereby obtaining a military simulation agent with the ability to resist irregular transient amplification.

2. The method for designing military simulation intelligent agents based on knowledge graphs according to claim 1, characterized in that, The coordinated action nodes and decision trigger condition nodes in the wargaming plan are instantiated as graph entities in the wargaming combat element graph, including: Iterate through all combat units, combat phases, and combat functions in the wargaming simulation plan. For each triplet combination consisting of a combat unit, combat phase, and combat function, generate a corresponding collaborative action node in the wargaming combat element graph, and establish graph association edges pointing to the corresponding combat unit, corresponding combat phase, and corresponding combat function for each collaborative action node. It also iterate through all decision points and all triggering conditions associated with decision points in the wargaming simulation plan. For each binary combination consisting of a decision point and a triggering condition, generate a corresponding decision triggering condition node in the wargaming combat element graph, and establish graph association edges pointing to the corresponding decision point and corresponding triggering condition for each decision triggering condition node.

3. The method for designing a military simulation intelligent agent based on knowledge graphs according to claim 2, characterized in that, Based on the topological connections of the wargaming combat element map, a multi-dimensional battlefield situation vector is generated, including: Define a fixed order for the expansion of the map nodes, extract the execution intensity values ​​of all cooperative action nodes in the wargaming combat element map, and arrange them to form a cooperative action intensity vector; extract the condition satisfaction values ​​of all decision trigger condition nodes in the wargaming combat element map, and arrange them to form a trigger condition satisfaction vector; vertically concatenate the cooperative action intensity vector and the trigger condition satisfaction vector to generate the multidimensional battlefield situation state vector; the total number of dimensions of the multidimensional battlefield situation state vector is equal to the sum of the total number of cooperative action nodes and the total number of decision trigger condition nodes.

4. The method for designing a military simulation intelligent agent based on a knowledge graph according to claim 1, characterized in that, Establish a linear state transition matrix representing the operational phase progression logic, the sequential coordination logic of combat functions, and the decision-triggered injection relationship, including: Based on the phase sequence definition in the wargaming plan, a phase advancement matrix representing the time dimension is constructed, which represents the logic of the phase evolution in a predetermined order. Based on the dependency order between different combat functions within the same phase in the wargaming plan, a combat function serial coordination matrix representing the logical dimension is constructed, which represents the directed serial coupling relationship of combat functions from preceding functions to subsequent functions. Based on the path search results in the wargaming combat element map, a decision trigger injection matrix representing the causal dimension is constructed. If there is an effective influence path from the decision trigger condition node to the coordinated action node in the wargaming combat element map, it is marked as an effective influence value at the corresponding position in the decision trigger injection matrix.

5. The method for designing a military simulation intelligent agent based on a knowledge graph according to claim 4, characterized in that, The linear state transition matrix constitutes the computational kernel of the combat situation evolution transmission model, including: Using Kronecker product operations, the combat unit identity matrix, the combat phase advancement matrix, and the combat function serial coordination matrix are coupled by tensor product to form a dominant advancement term that characterizes the natural evolution of the multidimensional battlefield situation under interference-free conditions. Using block matrix splicing operations, the decision trigger injection matrix is ​​multiplied by a fixed scaling constant and then superimposed as a bias block onto the upper right block region of the dominant advancement term, thereby forming the complete linear state transition matrix. The linear state transition matrix has a block-based triangular structure and super-diagonal coupling characteristics, which are used to characterize the discontinuous situational jumps and transient evolution processes caused by decision point triggers during wargaming.

6. The method for designing military simulation intelligent agents based on knowledge graphs according to claim 5, characterized in that, The dominant divergent risk mode is mapped to sensitivity and triggering attributes, and the sensitivity and triggering attributes are written back to the corresponding map entities in the wargaming combat element map to obtain a wargaming combat element map containing stability constraint attributes, including: Write the value of the sensitivity attribute into each corresponding collaborative action node in the wargaming combat element map; Write the value of the triggering attribute into each corresponding decision triggering condition node in the wargaming combat element map; The sensitivity attribute represents the risk level of the corresponding collaborative action being erroneously amplified in the simulation, and the triggering attribute represents the potential intensity of the divergence of the overall inference results induced by the decision triggering conditions.

7. The method for designing military simulation intelligent agents based on knowledge graphs according to claim 1, characterized in that, A deterministic situational awareness sequence is constructed based on the numerical ranking of the sensitivity attribute and the triggering attribute in the wargaming combat element map containing stability constraint attributes, including: All decision trigger condition nodes in the wargaming combat element map containing stability constraint attributes are extracted and sorted in descending order according to the triggering attribute value recorded by each decision trigger condition node to generate an ordered trigger condition state subsequence. All cooperative action nodes in the wargaming combat element map containing stability constraint attributes are extracted and sorted in descending order according to the sensitivity attribute value recorded by each cooperative action node to generate an ordered cooperative action state subsequence. The trigger condition state subsequence and the cooperative action state subsequence are concatenated in a fixed sequence to form the deterministic situational awareness sequence, thereby ensuring that the agent prioritizes the perception of battlefield situational information with high divergence risk.

8. The method for designing a military simulation intelligent agent based on a knowledge graph according to claim 7, characterized in that, By combining divergent suppression decision logic to generate agent action strategies, a military simulation agent with resistance to irregular transient amplification is obtained, including: Graph embedding technology is used to map each action entity in the wargaming simulation combat element graph into an embedding vector. For each candidate action, the corresponding cooperative action node and decision triggering condition node associated in the graph are retrieved. The sensitivity attribute value and triggering attribute value of the cooperative action node and decision triggering condition node are accumulated to calculate the divergence risk cost of the candidate action. The deterministic situational awareness sequence is mapped into a hidden state. The matching degree between the hidden state and the action embedding vector is calculated, and the weighted divergence risk cost is subtracted to obtain the final confidence of the candidate action. Based on the final confidence, the action with the largest value is selected as the agent's action strategy.

Images