Tactical simulation and deduction method based on heterogeneous computing power collaboration

Abstract

This invention discloses a tactical simulation and extrapolation method based on heterogeneous computing power collaboration, belonging to the field of tactical simulation and extrapolation technology. The method includes the following steps: First, by constructing a simulation index structure to distinguish between initialized and uninitialized units, newly added combat units are identified during simulation operation, and based on the identification results, these newly added combat units are marked as temporarily added combat units. Second, based on the marked temporarily added combat units, state change data and behavioral command sequences are collected within multiple consecutive simulation frames to construct a behavioral event sequence matrix. Third, a sliding window weighted summation algorithm is used to calculate the behavioral change amount between each frame, and a short-term behavioral change rate index is established on the behavioral event sequence matrix. This invention solves the problem of inaccurate computing power adaptation for temporarily added combat units, realizes dynamic adjustment of heterogeneous computing power resources based on behavioral change rate, and improves simulation response efficiency and timing accuracy.

Description

Technical Field

[0001] This invention relates to the field of tactical simulation and deduction technology, specifically to a tactical simulation and deduction method based on heterogeneous computing power collaboration. Background Technology

[0002] Tactical simulation based on heterogeneous computing power collaboration refers to the comprehensive utilization of various heterogeneous computing resources (such as CPUs, GPUs, FPGAs, etc.) in the process of modeling complex combat environments and simulating tactical behaviors, in order to improve the real-time performance, accuracy, and scalability of the simulation. Existing technologies of this kind typically construct a distributed simulation system architecture. First, the available computing resources are identified and tagged for management. Based on the computational density, parallelism requirements, and latency sensitivity of various simulation tasks, task partitioning and scheduling algorithms are used to allocate high-concurrency, high-load sub-tasks such as graphics rendering and physics simulation to parallel units such as GPUs, while serial tasks such as logic control and decision-making are handled by CPUs. When necessary, FPGAs or DSPs are used to handle low-latency, fixed-logic key modules, thereby achieving cross-platform computing power collaboration. The entire simulation process generally includes the following steps: modeling and initialization of the battlefield environment and combat units, scenario configuration and parameter setting of simulation tasks, scheduling and allocation of computing resources, execution and real-time interaction of multi-module parallel simulation calculations, data synchronization and status feedback, and dynamic display and evaluation of simulation results. Ultimately, through the efficient collaboration of heterogeneous computing power, the ability to rapidly predict the evolution of tactical behavior and perform multi-scenario simulations is achieved.

[0003] The existing technology has the following shortcomings:

[0004] During tactical simulations, when a new combat unit is temporarily added to the system, the scheduler typically completes the initial computational resource binding directly based on the unit's type information and the available heterogeneous computing resources in the system. Since this unit has not yet undergone actual behavior when added, the system can only statically estimate its computational requirements based on the unit's default computational power template, failing to accurately reflect its real-time behavioral characteristics during subsequent simulations. If the unit frequently triggers high-intensity tactical behaviors (such as rapid state switching, path replanning, and communication coordination) within a short period, its actual computational demand will far exceed the processing capacity of the bound resources, thus creating a computational bottleneck. However, existing tactical simulation techniques based on heterogeneous computing power collaboration cannot dynamically determine whether to adjust the bound heterogeneous computing resources based on the rate of behavioral change exhibited by the temporarily added combat unit within a short simulation period, thus failing to achieve adaptive optimization of resource allocation strategies. This will lead to problems such as delayed behavioral response and lagging state updates in the simulation, further affecting the timing accuracy of the overall simulation synchronization and the consistency of the evolution of tactical behavior, and reducing the system's processing efficiency and simulation credibility in highly dynamic task environments.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a tactical simulation and deduction method based on heterogeneous computing power collaboration to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a tactical simulation and deduction method based on heterogeneous computing power collaboration, specifically including the following steps:

[0008] S1. By constructing a simulation index structure to distinguish between initialized and non-initialized units, newly added combat units are identified during the simulation process, and based on the identification results, the newly added combat units are marked as temporarily added combat units.

[0009] S2. Based on the marked and temporarily added combat units, collect state change data and behavior command sequences within multiple consecutive simulation frames, construct a behavior event sequence matrix, and use a sliding window weighted summation algorithm to calculate the behavior change between each frame. Establish a short-term behavior change rate index on the behavior event sequence matrix.

[0010] S3. Input the short-term behavior change rate index into the resource adaptation assessment structure, and extract the resource response deviation value by establishing a multivariate mapping relationship between the behavior change rate and the response capability of the bound heterogeneous computing resources, and generate computing power adjustment candidate tags.

[0011] S4. Adjust candidate flags based on computing power, combine the computing power usage records and behavior change trends of temporarily added combat units in the first few simulation frames, construct a dynamic judgment logic path graph, and output the scheduling decision value of whether the heterogeneous computing power binding relationship needs to be adjusted through the condition judgment of the graph structure.

[0012] S5. Based on the scheduling decision value, perform an adaptive adjustment operation on the binding relationship between temporarily added combat units and heterogeneous computing resources, including when the scheduling decision value is in the rebinding state, unbinding the current binding and selecting heterogeneous computing resources that meet the resource adaptation conditions to establish a new binding relationship.

[0013] Preferably, S1 is as follows:

[0014] By constructing a simulation index structure to distinguish between initialized and uninitialized units, during the simulation initialization phase, a key-value pair consisting of a unit identifier and a simulation start frame number is generated for each initialized unit, and all key-value pairs are written into the simulation index structure to form a set of reference indexes for unit identification in subsequent simulation cycles.

[0015] During the simulation process, the unit identifier of the combat unit entering the simulation process and the current simulation frame number are extracted in real time to generate a comparison key-value pair. The comparison key-value pair is then matched with the benchmark index set. When no matching result exists and the current simulation frame number is greater than the frame number value in any benchmark key value, it is identified as a newly added combat unit.

[0016] Based on the identification results, the unit identifier and the current simulation frame number of the newly identified combat unit are written into the newly added unit record set, and a temporary addition mark is added to the unit identifier for subsequent processing of the behavior change rate of the temporarily added combat unit.

[0017] Preferably, S2 specifically includes the following steps:

[0018] S201. Based on the temporarily added combat units that have been marked, during the simulation operation, starting from the simulation frame in which the combat unit first appears, the state change data and behavior command sequence within multiple simulation frames are continuously collected, and the collected results are arranged into a time series set in order of frame number, and archived and stored by unit identifier for subsequent matrix construction.

[0019] S202. The state change data and behavior instruction sequence corresponding to each frame are uniformly formatted and encoded into behavior event vectors of the same dimension. All behavior event vectors are then filled into a two-dimensional structure with the unit identifier as the row index and the simulation frame number as the column index according to the time sequence to construct a behavior event sequence matrix.

[0020] S203. Set a sliding window structure on the behavior event sequence matrix, use a weighted summation algorithm to calculate the behavior change between adjacent frames for the behavior event vectors within each window range, and perform time normalization on the change value to form a short-term behavior change rate index, which is used to characterize the dynamic intensity of behavior within the window period.

[0021] Preferably, S203 is as follows:

[0022] A sliding window structure is constructed on the behavior event sequence matrix at fixed frame intervals. The width of the sliding window is equal to the number of consecutive simulation frames. The window slides linearly on the behavior event sequence matrix with a step size of one frame. In each slide, the set of behavior event vectors corresponding to all simulation frames within the current window range is extracted as the input basis for behavior change analysis.

[0023] For each sliding window, a vector interpolation operation is performed on the behavior event vectors between adjacent frames in the set of behavior event vectors. A weighted summation algorithm is then performed based on the preset behavior dimension weight configuration to calculate the behavior change between each frame within the window range and generate a sequence of behavior change as an intermediate output.

[0024] Divide each item in the behavior change sequence by the corresponding inter-frame time interval to obtain the behavior change intensity value per unit time. Then, perform window averaging on the behavior change intensity value sequence to generate a short-term behavior change rate index, which is used to characterize the dynamic intensity change trend of the behavior of temporarily added combat units within the current sliding window period.

[0025] Preferably, S3 is as follows:

[0026] The short-term behavior change rate index is paired with the heterogeneous computing resource identifiers bound to the temporarily added combat units to construct a unit behavior-computing resource input structure, which is then input into the resource adaptation evaluation structure. The unit identifier is used as the primary key index to ensure that the input data of each unit in the evaluation structure is independent and continuous.

[0027] In the resource adaptation assessment structure, the multivariate mapping construction function is called to unify and normalize the short-term behavior change rate index and the response capability parameters of heterogeneous computing resources, generate a multivariate mapping relationship model to reflect the correlation between the actual load of resources and the unit behavior rate, and calculate the difference between the unit current behavior change rate and the computing power response capability as the resource response deviation value based on the multivariate mapping relationship model.

[0028] The resource response deviation value is compared with the preset deviation threshold range. When the resource response deviation value exceeds the upper limit of the preset deviation threshold range, the current unit is marked as a candidate unit for computing power adjustment, and a computing power adjustment candidate tag containing the unit identifier and deviation value is generated for subsequent judgment and control of resource binding relationships.

[0029] Preferably, S4 specifically includes the following steps:

[0030] S401. Based on the unit identifier in the candidate marker for adjusting computing power, extract the computing power usage records and behavior change trends of the corresponding temporarily added combat units in the first few simulation frames, align the two time series by frame number, and construct a joint feature array with frame number as the main index and behavior rate and computing power usage as joint fields, as the basic input for subsequent path graph construction.

[0031] S402. Construct a dynamic judgment logic path graph based on the joint feature array. In the graph structure, the frame number is used as the path node, and the relationship pattern between the behavior change trend and the computing power usage record is used as the edge weight attribute. The nodes are connected in sequence to form the graph path structure. In the path structure, a judgment condition label is added to each side to express the association condition between the behavior surge and the computing power load mismatch.

[0032] S403. Traverse the judgment condition labels in the dynamic judgment logic path graph, identify path segments that continuously meet the abnormal edge weight conditions, calculate the score based on the path length, the number of consecutive triggers and the edge weight strength, and generate a scheduling decision value when the score value exceeds the adjustment judgment threshold. This value is used to output the judgment conclusion on whether the heterogeneous computing power binding relationship needs to be adjusted.

[0033] Preferably, S403 is as follows:

[0034] In the dynamic judgment logic path graph, all path nodes are traversed in chronological order, the judgment condition labels on each path are extracted, and for each edge's judgment condition label and edge weight parameter, edge elements that meet the abnormal relationship threshold are selected and their corresponding frame numbers are recorded as potential abnormal marker points.

[0035] Based on the marked potential anomaly markers, identify the continuously connected abnormal edge segments in the path graph, calculate the path length, number of consecutive triggers of the abnormal condition, and corresponding edge weight strength of each abnormal path segment, and construct a scoring function model by weighted combination based on logistic regression to generate a path score value that reflects the level of resource demand pressure.

[0036] The path score is compared with the adjustment judgment threshold. If the path score is higher than the adjustment judgment threshold, a scheduling decision value is generated for the temporarily added combat unit associated with the corresponding path segment. The judgment conclusion that the heterogeneous computing power binding relationship of the unit needs to be adjusted is output, and the scheduling state transition is triggered.

[0037] Preferably, S5 is as follows:

[0038] The scheduling decision value determines whether the temporarily added combat unit is in a rebinding state. When the scheduling decision value indicates that it is in a rebinding state, the temporarily added combat unit that is currently in a rebinding state is selected as the target unit, and the adaptive adjustment operation of the binding relationship is prepared to be executed.

[0039] Remove the binding relationship between the target unit and the currently bound heterogeneous computing resources, delete the resource identifier of the heterogeneous computing resources from the unit-resource mapping table, and write the resource identifier into the resource idle tag set to mark it as allocable;

[0040] Search the resource idle marker set for heterogeneous computing resources that meet the resource adaptation conditions. For each candidate resource, calculate the adaptation degree based on its response capability, historical load level and the short-term behavior change rate index of the target unit. Select the heterogeneous computing resource with the highest adaptation degree as the new binding target, update the unit-resource mapping table, and complete the adaptive adjustment operation of the binding relationship between the temporarily added combat unit and the heterogeneous computing resource.

[0041] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0042] 1. This invention achieves precise adaptation of computing resources for combat units temporarily added during tactical simulation by constructing a heterogeneous computing power binding mechanism with dynamic scheduling capabilities. The scheme uses short-term behavior change rate as the core quantitative feature, combined with behavior event sequence matrix construction and sliding window weighted calculation methods, to extract the dynamic intensity of unit behavior within consecutive simulation frames in real time. This effectively captures the surge trend of behavior and overcomes the problem of large errors in traditional static computing power prediction. Simultaneously, by leveraging a multivariate mapping relationship model established through a resource adaptation evaluation structure, the system can quantify the deviation between the unit's behavior rate and the response capability of the currently bound heterogeneous resources, generating computing power adjustment candidate markers with controllable accuracy, providing a precise basis for subsequent scheduling decisions.

[0043] 2. This invention introduces a dynamic judgment logic path graph based on time series analysis. Through conditional correlation analysis of path nodes and edge weight attributes, it establishes an anomaly identification mechanism between behavioral changes and computing power usage. Combined with a logistic regression scoring function, it achieves intelligent assessment of potential resource bottleneck states. When the scoring result exceeds the adjustment judgment threshold, the system automatically generates a scheduling decision value and triggers adaptive adjustment of resource binding relationships. This process integrates multi-dimensional factors such as computing power usage history, dynamic behavioral trends, and resource adaptability assessment, enabling resource binding relationships to have portability and elastic adjustment capabilities. This significantly improves the task processing efficiency and response stability of the simulation system in highly dynamic environments, effectively ensuring the temporal continuity and computational accuracy of simulation behavior. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0045] Figure 1 This is a flowchart illustrating the tactical simulation and deduction method based on heterogeneous computing power collaboration of the present invention. Detailed Implementation

[0046] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0047] This invention provides, for example Figure 1The tactical simulation and deduction method based on heterogeneous computing power collaboration shown includes the following steps:

[0048] S1. By constructing a simulation index structure to distinguish between initialized and non-initialized units, newly added combat units are identified during the simulation process, and based on the identification results, the newly added combat units are marked as temporarily added combat units.

[0049] In this embodiment, S1 specifically refers to:

[0050] By constructing a simulation index structure to distinguish between initialized and uninitialized units, during the simulation initialization phase, a key-value pair consisting of a unit identifier and a simulation start frame number is generated for each initialized unit, and all key-value pairs are written into the simulation index structure to form a set of reference indexes for unit identification in subsequent simulation cycles.

[0051] During the simulation initialization phase, a set of key-value pairs, "unit identifier - frame number," can be constructed by assigning a unique unit identifier to each initialization unit and recording the simulation start frame number corresponding to its entry into the simulation system. This set can be implemented using a hash table, a bidirectional mapping table, or a two-dimensional array to support efficient data insertion and lookup operations. In the specific implementation, the unit configuration data at the start of the simulation can be extracted by the initialization configuration file parser, and an identifier code can be automatically generated for each unit, for example, based on a type prefix plus a number sequence. Then, the start frame number of the current simulation timer is read and bound to the unit identifier, written as a key-value pair into the data structure. Once constructed, this set can serve as the basis for identifying the source of units during simulation operation. When a new unit identifier appears in a subsequent simulation frame, the system can query this set to determine whether the unit was registered during the initialization phase, thus laying the foundation for subsequent unit classification and labeling.

[0052] The simulation index structure, used to distinguish between initialized and non-initialized units, is a static data set built upon the simulation startup time. Its core technical feature lies in recording the identity of each initialized unit and its reference time point when it first participated in the simulation. This index structure serves not only as a static record of unit identity but, more importantly, as the ability to determine the "generation sequence" in the unit's time domain. The frame numbers of initialized units are typically concentrated at the beginning of the simulation timeline, while non-initialized units, i.e., units dynamically added during the simulation, usually have their first frame number appearing much later than their initial distribution. Therefore, comparing the timestamps based on the frame numbers becomes a key indicator for distinguishing between the two types of units. Through the joint storage and discrimination of unit identifiers and frame numbers, this simulation index structure can effectively support subsequent high-level tasks such as dynamic unit identification, behavior observation triggering, and computing power allocation filtering, making it an indispensable foundational component in simulation scheduling logic.

[0053] During the simulation process, the unit identifier of the combat unit entering the simulation process and the current simulation frame number are extracted in real time to generate a comparison key-value pair. The comparison key-value pair is then matched with the benchmark index set. When no matching result exists and the current simulation frame number is greater than the frame number value in any benchmark key value, it is identified as a newly added combat unit.

[0054] During simulation, for each combat unit entering the simulation process in each frame, the combination of its unit identifier and the current simulation frame number needs to be extracted, and these two elements are combined into a unique key-value pair. This key-value pair serves as the real-time identification input and is used to accurately match the simulation index set established during the initialization phase. The matching operation can achieve rapid location through hash mapping, binary search, or bitmap indexing mechanisms. If the query result shows that the unit identifier is not registered in the simulation index set, and the current frame number is significantly greater than the simulation starting frame number of any unit in the index set, then the unit can be confirmed as a newly added combat unit during the simulation. This implementation logic can avoid repeated identification and misjudgment of initial units, while ensuring the correct identification position of dynamically added units in the simulation time series, providing high timeliness and accuracy support for subsequent behavior analysis and computing power allocation.

[0055] Unit identifiers must be globally unique, typically generated using a unified encoding rule, such as a combination of task type prefix and timestamp sequence number. Simulation frame numbers are automatically advanced, driven by the simulation master clock, ensuring the comparability of the time sequence in which all units enter the system. The generation process of comparison key-value pairs requires unit identifiers and frame numbers to be bound in a structured manner, facilitating efficient table lookup logic. The key to the matching process lies in a dual-condition judgment mechanism: not only must the unit identifier not exist in the index set, but the current frame number must also be later than the initial distribution to accurately locate the newly added unit. This dual-condition filtering mechanism avoids re-identification of units during the initialization phase and enhances the time sensitivity and logical rigor of new unit identification.

[0056] Based on the identification results, the unit identifier and the current simulation frame number of the newly identified combat unit are written into the newly added unit record set, and a temporary addition mark is added to the unit identifier for subsequent processing of the behavior change rate of the temporarily added combat unit.

[0057] After identifying newly added combat units, the identified unit identifiers and their corresponding current simulation frame numbers need to be written into a dedicated dataset to record all units dynamically added during the simulation. This dataset can be implemented using a scalable hash map structure or a linked list structure arranged in time series to support subsequent unit retrieval and time positioning. The writing process should ensure the uniqueness of the unit identifier and bind it with the current frame number as a time reference to determine the starting position of its behavior change in subsequent behavior calculations. Simultaneously, a special logical flag is added to the unit identifier to indicate that it is a temporarily added combat unit. This flag field can be used as an input filtering condition for the behavior change rate analysis function, executing behavior tracking logic only on units marked with this flag. This isolates the processing flow of initialization units from that of runtime units at the data processing level, improving system efficiency and data structure organization.

[0058] The core of this process is concentrated in three aspects: First, the dual-element binding method of unit identification and time information ensures that the record set has both identity indexing function and time series reference capability; second, the data structure selection of the record set must support fast lookup, batch update, and time sorting, with commonly used structures including time hash tables with indexed tables or ordered linked lists driven by simulation frames; third, the generation and binding mechanism of temporary added tags, which should have unique Boolean state or level identification attributes and be directly attached to the unit identification through binding relationships, serving as the basis for execution branch determination during the behavior analysis phase. Through the above mechanisms, newly added units can be logically and effectively distinguished from all combat units, providing an accurate and efficient data foundation for subsequent calculation processes such as behavior change rate extraction and computing resource adaptation.

[0059] S2. Based on the marked and temporarily added combat units, collect state change data and behavior command sequences within multiple consecutive simulation frames, construct a behavior event sequence matrix, and use a sliding window weighted summation algorithm to calculate the behavior change between each frame. Establish a short-term behavior change rate index on the behavior event sequence matrix.

[0060] In this embodiment, S2 specifically includes the following steps:

[0061] S201. Based on the temporarily added combat units that have been marked, during the simulation operation, starting from the simulation frame in which the combat unit first appears, the state change data and behavior command sequence within multiple simulation frames are continuously collected, and the collected results are arranged into a time series set in order of frame number, and archived and stored by unit identifier for subsequent matrix construction.

[0062] During simulation, when a combat unit enters a simulation frame, its unique identifier and current frame number are monitored in real time. To dynamically capture the behavior of newly added units, it is necessary to continuously record the unit's state change data and behavioral command sequences across multiple simulation frames, starting from the frame in which the unit first appears. State change data includes dynamic attributes such as spatial position, velocity vector, and attitude information, while behavioral command sequences contain logical information such as decision-making behavior type, response trigger type, and action execution flags. The data acquisition method can be based on a simulation memory callback mechanism. At each frame refresh, the attributes and behaviors of the target unit are extracted using a unified data interface, and a time series set is constructed according to the frame number. This time series set is then written into the corresponding data channel based on the unit identifier to form a structured archive. This processing method ensures a stable time axis index and consistent data dimensions when constructing the behavioral event sequence matrix, facilitating subsequent analysis of behavioral change trends.

[0063] "Temporarily added combat units that have been marked" refers to units dynamically inserted during the simulation run and whose identities have been confirmed through the simulation index structure and matching mechanism. Continuously collecting data from multiple simulation frames is to construct a behavioral data sequence that reflects the evolutionary patterns over time. It is generally recommended that the number of frames cover the unit's complete behavioral cycle within a short period to ensure feature integrity. State change data and behavioral command sequences are two key dimensions of the unit's behavioral logic and resource demand changes. The former reflects physical dynamics, and the latter reflects logical responses; combining them can reconstruct the unit's complete behavioral trajectory within the simulation cycle. Assembling the collected results into a time series set in frame number order ensures the consistency and retrieval of the data structure over time. Archiving and storing data by unit identifier provides a unique index basis for subsequently constructing a behavioral event sequence matrix by unit dimension, thereby ensuring the accuracy and controllability of behavioral change extraction and velocity index construction.

[0064] S202. The state change data and behavior instruction sequence corresponding to each frame are uniformly formatted and encoded into behavior event vectors of the same dimension. All behavior event vectors are then filled into a two-dimensional structure with the unit identifier as the row index and the simulation frame number as the column index according to the time sequence to construct a behavior event sequence matrix.

[0065] To align state change data and behavioral command sequences across different simulation frames in the time dimension, these two types of data need to be uniformly formatted and encoded. Specifically, a preset data standard template can be used to map various parameters (such as position, velocity, and attitude) in the state change data and various flags (such as command type, priority, and response trigger) in the behavioral command sequence into fixed-dimensional numerical vectors, with each dimension representing a behavioral feature. After encoding, the behavioral event vectors generated for each simulation frame are stored sequentially in a two-dimensional array structure, where the row index is the unit identifier and the column index is the simulation frame number. This ultimately forms a behavioral event sequence matrix with the unit as the main axis, the time frame as the sequence, and a unified data dimension. This structure not only solves the problem of aligning heterogeneous data but also provides a basic structural carrier for subsequent sliding window processing and behavioral change calculation, significantly improving the quantifiable processing capability of behavioral dynamics.

[0066] State change data and behavioral command sequences originate from the simulation engine's continuous monitoring of combat units. Due to their inconsistent dimensions and complex content, they must be structurally fused through unified format encoding. The core feature of behavioral event vectors lies in their homogeneity; each encoded vector represents a frame of behavioral data in time and maintains consistent spatial dimensions, facilitating matrix operations. In the two-dimensional structure, using unit identifiers as row indices enables isolated processing between units, avoiding cross-contamination of behavioral features. Using simulation frame numbers as column indices ensures temporal continuity and comparative consistency, guaranteeing that behavioral change trends can be fully captured by the sliding window. This behavioral event sequence matrix is ​​a crucial bridge for abstracting temporal behavioral information from raw data into structured information, serving as the core foundational data structure for subsequent sliding statistics, trend identification, and computational power assessment. The overall design reflects a high degree of attention to data computability and structural stability, providing pre-processing data support for dynamic scheduling decisions of heterogeneous computing resources.

[0067] S203. Set a sliding window structure on the behavior event sequence matrix, use a weighted summation algorithm to calculate the behavior change between adjacent frames for the behavior event vectors within each window range, and perform time normalization on the change value to form a short-term behavior change rate index, which is used to characterize the dynamic intensity of behavior within the window period.

[0068] A sliding window structure is set on the behavior event sequence matrix, and a weighted summation algorithm is used to calculate the behavior change between adjacent frames within each window. The change values ​​are then time-normalized to form a short-term behavior change rate index. The core purpose is to accurately characterize the dynamic intensity changes in the behavior of temporarily added combat units within a specific time interval. Since the behavior characteristics of newly added units are not yet stable, directly using static templates is insufficient to reflect their real-time behavior load characteristics. The sliding window mechanism enables continuous monitoring within a local time domain, thereby dynamically capturing the rhythm of unit behavior fluctuations in the short term. The weighted summation algorithm further enhances the sensitivity to key behavior dimensions, highlighting the impact of high-frequency or high-intensity behaviors through difference operations and weight fusion. Time normalization ensures the consistency of the calculation results across time scales, effectively eliminating the interference of different simulation frame rates or sampling densities on behavior evaluation. Therefore, this approach can quickly and objectively establish a behavior load profile without relying on long-term historical data, providing a reasonable quantitative basis and dynamic judgment foundation for subsequent rebinding of heterogeneous computing resources.

[0069] In this embodiment, S203 specifically refers to:

[0070] A sliding window structure is constructed on the behavior event sequence matrix at fixed frame intervals. The width of the sliding window is equal to the number of consecutive simulation frames. The window slides linearly on the behavior event sequence matrix with a step size of one frame. In each slide, the set of behavior event vectors corresponding to all simulation frames within the current window range is extracted as the input basis for behavior change analysis.

[0071] The purpose of setting a sliding window structure in the behavior event sequence matrix is ​​to capture the continuity and abrupt changes in the short-term behavior of combat units on the time axis. The sliding window structure divides time intervals on the matrix with a fixed frame width, and performs linear sliding in one-frame increments, ensuring frame-by-frame coverage of the behavior event vector sequence. In implementation, the width parameter of the sliding window can be configured, typically determined based on the simulation engine's refresh rate and behavior response cycle, ensuring that each window cycle covers the complete behavioral characteristics of the unit within a short time. During each sliding operation, the set of behavior event vectors from all frames within the current window interval is extracted and used as input for behavior change analysis. The sliding window structure is essentially a temporal analysis framework that not only preserves the temporal continuity of behavioral data but also provides contextual support for subsequent change intensity extraction through data aggregation within the window. In this way, local analysis of the stability and dynamics of unit behavior in different time intervals can be achieved, thus enabling high-resolution behavior trend modeling and response monitoring without disrupting the overall simulation structure. The key to the design of the sliding window structure lies in the coordinated setting of the inter-frame step size and the window width, which ensures both computational accuracy and avoids redundant computation, effectively improving the real-time performance and sensitivity of behavior fluctuation recognition.

[0072] For each sliding window, a vector interpolation operation is performed on the behavior event vectors between adjacent frames in the set of behavior event vectors. A weighted summation algorithm is then performed based on the preset behavior dimension weight configuration to calculate the behavior change between each frame within the window range and generate a sequence of behavior change as an intermediate output.

[0073] Within each sliding window, to measure the intensity of behavioral changes in combat units across consecutive simulation frames, a vector interpolation operation needs to be performed on the set of behavioral event vectors. Specifically, for each pair of adjacent simulation frames arranged chronologically within the window, their corresponding behavioral event vectors are extracted, and element-wise subtraction is performed between them to form a difference vector, reflecting the magnitude of changes in each behavioral dimension between the unit and the frame pair. Subsequently, a set of preset behavioral dimension weight configurations is introduced, with each behavioral dimension corresponding to a weight coefficient. This weight value is set based on the importance of the behavioral type or computational consumption characteristics; for example, the weight of the state transition dimension can be higher than that of the path point update dimension. After performing element-wise multiplication of the difference vector and the behavioral dimension weight vector, the resulting vector is summed to obtain the weighted total change value between the frame pair, which is the behavioral change amount. In this way, the original multidimensional behavioral changes can be transformed into a single numerical expression, forming a sequence of behavioral change amounts that is easy to analyze and process. To illustrate with a concrete example, if a unit's behavior within a window spans five frames, including frequent turning, speed adjustments, and target locking, and each behavioral dimension is assigned a different weight value, the system can accurately calculate the degree of fluctuation in the unit's behavior across consecutive frames using a weighted difference method based on the intensity of changes in these dimensions. This method not only considers the frequency of behavior occurrences but also introduces a quantitative trade-off of dimensional importance, resulting in a more expressive and discriminative sequence of behavioral changes. This provides a mathematical foundation and behavioral basis for the next step of generating short-term behavioral change rate indicators.

[0074] Divide each item in the behavior change sequence by the corresponding inter-frame time interval to obtain the behavior change intensity value per unit time. Then, perform window averaging on the behavior change intensity value sequence to generate a short-term behavior change rate index, which is used to characterize the dynamic intensity change trend of the behavior of temporarily added combat units within the current sliding window period.

[0075] After extracting the behavioral change sequence, to assess the activity level of a unit within a sliding window period, each behavioral change needs to be divided by its corresponding inter-frame time interval to obtain the behavioral change intensity value per unit time, i.e., the degree of behavioral fluctuation per second. This calculation method ensures that the behavioral intensity assessment has time-standardized characteristics, making behavioral changes at different time steps comparable. For example, within a certain sliding window, if a unit generates a large behavioral change between two frames, but the inter-frame time is long, the change intensity value will be diluted; conversely, if the behavioral change is concentrated in a short period, the change intensity value will be significantly increased. To further stabilize the intensity assessment results and eliminate the random influence of local fluctuations, all behavioral change intensity values ​​per unit time are combined into an intensity sequence, and window averaging is performed, i.e., averaging all values ​​within the sequence, ultimately generating a short-term behavioral change rate index. This index reflects the overall dynamic activity level of the unit within the current sliding window period, possessing good behavioral pattern representation ability and evolutionary trend sensitivity. Taking a combat unit rapidly switching between multiple commands within a single window period as an example, its corresponding intensity value will remain consistently high, and the average result will also rise significantly. This indicates that the unit is currently in a high-behavioral-density phase, making it suitable as a trigger for resource scheduling optimization. In this way, the short-term behavioral fluctuations of units can be characterized in a refined manner, providing accurate and quantitative input for subsequent resource adjustment decisions.

[0076] S3. Input the short-term behavior change rate index into the resource adaptation assessment structure, and extract the resource response deviation value by establishing a multivariate mapping relationship between the behavior change rate and the response capability of the bound heterogeneous computing resources, and generate computing power adjustment candidate tags.

[0077] In this embodiment, S3 specifically refers to:

[0078] The short-term behavior change rate index is paired with the heterogeneous computing resource identifiers bound to the temporarily added combat units to construct a unit behavior-computing resource input structure, which is then input into the resource adaptation evaluation structure. The unit identifier is used as the primary key index to ensure that the input data of each unit in the evaluation structure is independent and continuous.

[0079] Pairing short-term behavior change rate indicators with heterogeneous computing resource identifiers bound to temporarily added combat units can be achieved by establishing a key-value mapping table. Specifically, during the simulation cycle, a unique identifier for each temporarily added combat unit and the identifier of its bound computing resource are maintained. By traversing the behavior change rate indicator queue, this is combined with the computing resource identifier to form an input structure. This input structure can be constructed as a multi-field data structure containing unit identifier, behavior rate value, computing resource type, and current resource load status, and is input as a unit data row into the resource adaptation evaluation structure. Using the unit identifier as the primary key index ensures that the evaluation data corresponding to each unit is stored independently and updated in real time in the data table, preventing data cross-contamination between units and improving evaluation accuracy and structure retrieval efficiency. This structure enables the evaluation mechanism to perform parallel analysis of resource adaptation status at the unit level, meeting the needs of dynamic high-concurrency evaluation.

[0080] The Unit Behavior-Computing Resource Input Structure is a data structure reflecting the coupling relationship between the current behavioral characteristics of combat units and the computing resources they rely on. Its core lies in binding short-term behavioral change rate indicators with heterogeneous computing resource identifiers and the current operating load status of the resources. This structure uses standardized field definitions to store multi-source heterogeneous data in a unified format, facilitating subsequent multi-dimensional analysis and judgment. The Resource Adaptation Evaluation Structure is a data processing structure used to evaluate the matching degree between unit behavioral rate and computing power response capability. It typically includes an input receiving component, a multivariate correlation modeling module, a difference calculator, and an evaluation result output cache. This structure continuously and quantitatively evaluates the rationality of resource allocation and the adaptability of the current load by iteratively inputting behavioral rate and computing power information frame by frame, serving as a fundamental component for dynamic resource judgment and rescheduling. Through the collaborative construction of these two structures, the system can achieve dynamic diagnostic capabilities for resource allocation status based on real-time behavioral data.

[0081] In the resource adaptation assessment structure, the multivariate mapping construction function is called to unify and normalize the short-term behavior change rate index and the response capability parameters of heterogeneous computing resources, generate a multivariate mapping relationship model to reflect the correlation between the actual load of resources and the unit behavior rate, and calculate the difference between the unit current behavior change rate and the computing power response capability as the resource response deviation value based on the multivariate mapping relationship model.

[0082] In the resource adaptation assessment structure, calling the multivariate mapping construction function allows for the quantitative modeling of the relationship between behavioral speed and resource response capability through multi-input regression modeling or shallow neural networks. Specifically, the process involves first standardizing the short-term behavioral change rate index and the heterogeneous computing resource response capability parameters to the same number of dimensions, avoiding modeling errors caused by inconsistent inputs. Next, normalization methods (such as Z-score or Min-Max normalization) are used to adjust input values ​​from different sources to the same dimension, ensuring that data from different units and scales can participate in unified calculations. Then, the preprocessed input vector is fed into a trained multivariate mapping model, which can be a function fitting structure trained on a large amount of historical simulation data, outputting the difference between the current behavioral speed and the computing power response. This difference value is the resource response deviation value; a larger value indicates that the currently bound resource cannot respond promptly to changes in the unit's behavioral intensity. For example, if a unit's current behavioral rate is as high as 20 state updates per second, while its bound resource's maximum processing frequency only supports 12 updates per second, the difference value is positive, requiring further adjustment and judgment.

[0083] A multivariate mapping construction function is a function mapping body used to establish a high-dimensional correlation between unit behavior indicators and computational resource response capabilities. Its function is to jointly model multiple input dimensions and output a single metric reflecting the resource matching degree of the current pairing state. This function typically includes an input standardization interface, a weight fusion layer, a mapping generation engine, and an error output module. The input standardization interface unifies the input range and format from different data sources, while the weight fusion layer automatically assigns weight values ​​to each input dimension based on historical training samples, ensuring that important dimensions have a stronger impact on the output. The mapping generation engine is responsible for performing nonlinear combination or kernel function transformations to extract key interaction features between behavior rates and resource parameters. Finally, the error output module generates and outputs the resource response deviation value. By constructing this function, accurate evaluation of the relationship between complex behaviors and resource capabilities can be achieved, supporting subsequent scheduling optimization decisions and avoiding scheduling mismatch problems caused by static rules.

[0084] The resource response deviation value is compared with the preset deviation threshold range. When the resource response deviation value exceeds the upper limit of the preset deviation threshold range, the current unit is marked as a candidate unit for computing power adjustment, and a computing power adjustment candidate tag containing the unit identifier and deviation value is generated for subsequent judgment and control of resource binding relationships.

[0085] The process of comparing resource response deviation values ​​with preset deviation threshold ranges can be achieved by setting upper and lower boundary values. The deviation value represents the difference between the unit's current behavior rate and the response capability of the bound heterogeneous computing resources. If this value exceeds the acceptable error range, it means that the currently bound resources cannot meet the unit's actual computational needs. In practice, a statically or dynamically generated deviation threshold range is first set; for example, a reasonable fluctuation range can be calculated based on the mean and standard deviation from the initial simulation. Each time a newly calculated resource response deviation value is generated, it is compared with the upper and lower boundaries of this range. If the deviation value is greater than the upper boundary value, the unit is determined to be a resource-insufficient risk unit, and the system will immediately add a computing power adjustment candidate mark for it. The computing power adjustment candidate mark consists of a unit identifier and the current deviation value, used to indicate that the unit has exceeded the current resource carrying capacity and needs to be prioritized in subsequent scheduling processes. For example, if a unit's deviation value is 0.38 and the preset upper limit is 0.30, because the deviation value exceeds the limit, the unit will be marked as a candidate unit, awaiting subsequent computing power rebinding judgment.

[0086] A preset deviation threshold range is a reference boundary range used to define the acceptable error range between resource response capability and unit behavior rate. Its definition methods typically include static configuration and dynamic learning. Static configuration can be predefined during the simulation platform deployment phase, for example, based on empirical threshold ranges derived from statistical analysis of past unit behavior fluctuation data of the same type. Dynamic learning, on the other hand, allows for online adjustment during simulation operation by continuously observing historical unit deviation performance, making the range more closely match the real-time simulation load environment. The threshold range typically consists of a minimum deviation value and a maximum deviation value, with the middle part representing the normal fluctuation range between unit behavior and resource capability. When the resource response deviation value is within this range, the system assumes normal resource matching and requires no adjustment; once it exceeds the upper limit, it means that the unit may experience behavioral lag or response congestion, thus requiring the initiation of a computing power adjustment process. Therefore, the preset deviation threshold range plays a core role in quantitative assessment in computing power matching judgment and is the foundation for implementing refined resource scheduling strategies.

[0087] S4. Adjust candidate flags based on computing power, combine the computing power usage records and behavior change trends of temporarily added combat units in the first few simulation frames, construct a dynamic judgment logic path graph, and output the scheduling decision value of whether the heterogeneous computing power binding relationship needs to be adjusted through the condition judgment of the graph structure.

[0088] In this embodiment, S4 specifically includes the following steps:

[0089] S401. Based on the unit identifier in the candidate marker for adjusting computing power, extract the computing power usage records and behavior change trends of the corresponding temporarily added combat units in the first few simulation frames, align the two time series by frame number, and construct a joint feature array with frame number as the main index and behavior rate and computing power usage as joint fields, as the basic input for subsequent path graph construction.

[0090] To achieve accurate judgment of scheduling status, a data structure is needed to characterize the coupling relationship between the dynamic behavior of combat units and the usage status of computing resources within a short period of time. The specific implementation is as follows: First, based on the unit identifier in the computing power adjustment candidate flags, the historical operational data of that unit within the simulation system is retrieved, including the computing power resource consumption index and behavior change rate index of the unit in each simulation frame. Computing power usage records can be collected by monitoring parameters such as the occupancy rate of heterogeneous computing resources bound to the unit in each frame, memory bandwidth usage, and the number of thread calls, while the behavior change trend is derived from the short-term behavior change rate index calculated in the preceding stage. Subsequently, the two time series are aligned along a unified time axis according to frame numbers to ensure a one-to-one correspondence between the computing power usage value and the behavior change rate index corresponding to the same frame number. Using the frame number as the main index, the two index values ​​of each frame are combined into a joint feature vector, constructing a two-dimensional array structure as the basic input for subsequent path graph node features. This array enables quantitative modeling of the behavior-resource relationship of units within continuous time periods, thus providing accurate support for the construction of edge weights and condition labels in the path graph.

[0091] Computing power usage records refer to the resource usage of heterogeneous computing resources during unit operation, including but not limited to a set of indicators reflecting unit load pressure such as floating-point operation ratio, core occupancy time, cache hit rate, and heterogeneous core switching count. These indicators need to be collected at the frame level and recorded in real time through the resource monitoring interface. Behavioral change trends refer to the rate of change of behavioral intensity within a short time window. This trend is generated by combining a behavioral event sequence matrix with sliding window weighting, and is usually expressed as the statistical slope or fluctuation range of a certain behavioral vector change. The joint feature array is a two-dimensional array with frame number as the primary index, where each item contains two fields: behavioral rate and computing power usage. This array achieves the fusion of behavioral and resource layer data in the time dimension. This structure can serve as the input baseline in graph neural networks or graph traversal logic, promoting the improvement of graph structure construction and edge weight assignment accuracy in dynamic judgment logic path graphs. It also facilitates time-series analysis and control judgment of the risk of resource mismatch caused by behavioral surges.

[0092] S402. Construct a dynamic judgment logic path graph based on the joint feature array. In the graph structure, the frame number is used as the path node, and the relationship pattern between the behavior change trend and the computing power usage record is used as the edge weight attribute. The nodes are connected in sequence to form the graph path structure. In the path structure, a judgment condition label is added to each side to express the association condition between the behavior surge and the computing power load mismatch.

[0093] The purpose of constructing a dynamic judgment logic path graph is to establish a time-series driven graph structure to characterize the temporal evolution relationship between the behavioral change trends of temporarily added combat units and computing power usage records within continuous simulation frames. The path graph is constructed based on a joint feature array, where each item corresponds to a simulation frame, containing behavioral change rate indicators and computing power resource usage data for that frame. The frame number in the joint feature array serves as the node identifier for the path graph, and nodes corresponding to adjacent frame numbers are connected sequentially in chronological order to form a linear directed graph structure. On the edges generated between any two adjacent nodes, the relationship between behavioral rate and computing power usage between the current frame and the previous frame is calculated. For example, if the behavioral rate increases but computing power usage does not increase significantly, this pattern is defined as an edge weight of the "computing power response lag" type. Edge weight attributes are assigned to edges according to different relationship patterns, and judgment condition labels are attached to each edge, such as "behavior surge with no resource increase" or "behavior stable but resource surge," to capture potential mismatches between behavior and resource responses. This graph structure can support subsequent abnormal pattern recognition and path scoring operations, providing a structured decision-making basis for dynamic scheduling.

[0094] The dynamic decision logic path graph is a graph structure built on frame-level time series, with three core technical features: path nodes, edge weight attributes, and decision condition labels. Path nodes, numbered by simulation frame, identify key state points in the simulation time series. Edge weight attributes describe the coupling relationship between behavioral change trends and computing power usage records between two adjacent nodes. Edge weights can be analyzed by constructing difference vectors to examine the relative relationship between behavioral increments and computing power increments, thus classifying them into response matching or response imbalance types. Decision condition labels are logical decision elements attached to the edges, indicating whether a behavior-resource pairing falls into a predefined set of abnormal decision patterns. These labels are generated based on joint triggering rules of behavioral change thresholds and resource usage thresholds, forming a traversable and quantifiable set of decision paths in the path graph. The dynamic decision logic path graph enables multi-dimensional modeling of the dynamic relationship between behavior and computing power, providing a visualized and scoreable structured foundation for scheduling decisions.

[0095] S403. Traverse the judgment condition labels in the dynamic judgment logic path graph, identify path segments that continuously meet the abnormal edge weight conditions, calculate the score based on the path length, the number of consecutive triggers and the edge weight strength, and generate a scheduling decision value when the score value exceeds the adjustment judgment threshold. This value is used to output the judgment conclusion on whether the heterogeneous computing power binding relationship needs to be adjusted.

[0096] By traversing the judgment condition labels in the dynamic judgment logic path graph and identifying path segments that continuously meet the abnormal edge weight conditions, the aim is to accurately capture the correlation between insufficient computing power response and surges in behavior of combat units within a specific time window in a highly dynamic tactical simulation environment, thereby improving the targeting and timeliness of heterogeneous computing power resource scheduling. By constructing a graph structure with frame numbers as nodes and behavior-computing power relationships as edge weights, the behavior and resource usage trajectory of units evolving over time can be transformed into an analyzable graph path pattern. Then, the abnormal risk level is comprehensively modeled using three elements: edge weight strength, path continuity, and trigger frequency. When the score exceeds the set adjustment judgment threshold, it indicates that the dynamic intensity of the unit's current behavior has continuously exceeded the tolerance of resource response capacity, belonging to a high-risk resource matching imbalance state. At this time, a scheduling decision value is generated, enabling the system to trigger computing power rebinding operations in a logically clear and quantifiable manner. This avoids the scheduling lag or over-adjustment problems caused by relying on static rules or single-index triggers in traditional scheduling strategies, fundamentally enhancing the system's rapid adaptability to sudden tactical demands and the operational stability of simulation tasks in a multi-source heterogeneous computing environment.

[0097] In this embodiment, S403 specifically refers to:

[0098] In the dynamic judgment logic path graph, all path nodes are traversed in chronological order, the judgment condition labels on each path are extracted, and for each edge's judgment condition label and edge weight parameter, edge elements that meet the abnormal relationship threshold are selected and their corresponding frame numbers are recorded as potential abnormal marker points.

[0099] To identify potential resource scheduling anomalies in simulations caused by temporarily added combat units, a traversal analysis of the dynamic judgment logic path graph is required. Specifically, a time-series traversal algorithm can be used, visiting all path nodes in the graph sequentially according to the simulation frame numbers. Each traversal involves the path edges between two adjacent nodes, extracting the attached judgment condition labels and edge weight parameters. Judgment condition labels typically describe the logical relationship between behavior rate and computing power consumption, such as "behavior rate increases sharply while resource load remains unchanged" or "resource usage increases but behavior changes tend to stabilize." Anomaly relationship thresholds are preset logical judgment rule boundaries, such as behavior changes exceeding a certain value but resource response falling below a certain proportion. These thresholds can be set based on historical simulation data statistics and dynamically configured for different tactical mission scenarios. During the traversal, by comparing whether the judgment condition labels of the current path edges meet the anomaly relationship thresholds, edge elements judged as having anomalies are filtered out, and the frame numbers of the successor nodes connected to these edges are recorded as potential anomaly markers for subsequent scoring and scheduling judgments. In this approach, the system can quickly detect potential resource allocation delays or overloads and provide accurate time-based information for scheduling decisions.

[0100] Based on the marked potential anomaly markers, identify the continuously connected abnormal edge segments in the path graph, calculate the path length, number of consecutive triggers of the abnormal condition, and corresponding edge weight strength of each abnormal path segment, and construct a scoring function model by weighted combination based on logistic regression to generate a path score value that reflects the level of resource demand pressure.

[0101] To quantify the resource adaptation pressure that temporarily added combat units may face within a specific time period, it is necessary to identify continuously connected abnormal edges from the path graph based on potential anomaly markers and to score each abnormal path segment. Specifically, all potential anomaly markers are first sorted chronologically, and then their connectivity in the graph structure is examined sequentially to identify path segments with consecutive frame numbers and edge condition labels that satisfy the anomaly relationship. For each identified abnormal path segment, its path length (representing the anomaly duration), the number of consecutive triggers of the anomaly condition (representing anomaly stability), and edge weight strength (reflecting the degree of imbalance between behavior and computing power) are calculated. These three features are then input into a weighted combined scoring model based on logistic regression for scoring. The weighted combination method of logistic regression here sets weight factors for multiple input features and combines them with a bias term to output a normalized score value. This score value reflects the degree of mismatch between the abnormal behavior and resource response of the unit corresponding to the current path segment, thus representing the pressure level at the resource scheduling level. For example, if a path segment has a long duration, frequent triggers, and extremely high edge weights, its score will be much higher than other segments. The logistic regression model is used instead of a simple weighted average because it can fit parameters to the combined effects of outliers based on historical data, exhibiting stronger generalization ability and assessment accuracy, and is applicable to various types of tactical simulation scenarios. This approach allows for a quantitative assessment of computing power scheduling risks, providing a precise basis for subsequent scheduling decisions.

[0102] The path score is compared with the adjustment judgment threshold. If the path score is higher than the adjustment judgment threshold, a scheduling decision value is generated for the temporarily added combat unit associated with the corresponding path segment. The judgment conclusion that the heterogeneous computing power binding relationship of the unit needs to be adjusted is output, and the scheduling state transition is triggered.

[0103] To determine whether temporarily added combat units require adjustments to their heterogeneous computing power binding relationships, the path score is compared with an adjustment threshold to make a scheduling decision. This process involves: first, setting a set of adjustment thresholds to distinguish between resource carrying capacity and scheduling triggering critical states. These thresholds are obtained through statistical learning of the dynamic intensity of unit behavior and resource response imbalances in historical simulation data, and have clear numerical boundaries. Then, the path score generated based on logistic regression in the previous step is compared with these thresholds one by one. If the score exceeds the threshold, it indicates a significant difference in response between the unit's current rate of change and its bound heterogeneous computing resources, suggesting a potential performance bottleneck. In this case, the system immediately generates a scheduling decision value for the unit and records it as a unit requiring computing power binding adjustments. The scheduling decision value can be represented by a status code indicating whether to enter the rescheduling process and submitted to the scheduling control module to perform resource unbinding and rebinding operations. Adjusting the threshold setting directly affects scheduling sensitivity and resource utilization efficiency. Its range design must fully consider the upper tolerance limit between the fluctuation range of unit behavior and resource adaptability. Typically, in large-scale, highly dynamic tactical simulations, a higher threshold strategy can effectively avoid frequent scheduling interference with simulation stability, while in resource-constrained scenarios, the threshold needs to be appropriately lowered to improve response sensitivity. Through this comparison process, clear determination and dynamic triggering of scheduling behavior can be achieved.

[0104] S5. Based on the scheduling decision value, perform an adaptive adjustment operation on the binding relationship between temporarily added combat units and heterogeneous computing resources, including when the scheduling decision value is in the rebinding state, unbinding the current binding and selecting heterogeneous computing resources that meet the resource adaptation conditions to establish a new binding relationship.

[0105] In this embodiment, S5 specifically refers to:

[0106] The scheduling decision value determines whether the temporarily added combat unit is in a rebinding state. When the scheduling decision value indicates that it is in a rebinding state, the temporarily added combat unit that is currently in a rebinding state is selected as the target unit, and the adaptive adjustment operation of the binding relationship is prepared to be executed.

[0107] In simulation systems, to dynamically manage the binding relationships between temporarily added combat units and heterogeneous computing resources, a scheduling decision value parser can be preset in the scheduling module to parse the current scheduling status flag of each unit. The scheduling decision value is organized by mapping the unit identifier as the key and the status value as the value. The status value can include different status types such as maintaining binding, rebinding, and prioritizing binding. When the parser determines that a unit's status is rebinding, the system extracts that unit from the current simulation running unit and adds it to the scheduling queue as a pending processing flag. This operation ensures that the system can respond promptly to resource mismatch issues and provides a clear scope for subsequent resource unbinding and reallocation. For example, if a unit's behavior rate consistently exceeds its resource processing capacity, it is marked by the evaluation structure as needing status adjustment. The scheduling module identifies its status as rebinding and can then enter the binding adjustment process, ensuring the real-time and targeted nature of dynamic resource allocation.

[0108] The scheduling decision value is a dynamically generated control signal, derived from the resource adaptation assessment structure's judgment of changes in unit computing power requirements. Temporarily added combat units refer to unit entities newly added during simulation operation, possessing non-initialized characteristics. The rebinding state is a judgment condition used to trigger resource unbinding and reallocation logic, typically output by path graph scoring or deviation value comparison processes. When the scheduling module parses the target unit's current state as rebinding, the system sets a status flag for that unit in the runtime data structure and prioritizes its processing in the resource scheduling flow, thereby achieving the orderly initiation of the resource reconfiguration process. This step provides clear preconditions for subsequent unbinding and rebinding operations and is the entry control point for the entire binding relationship adjustment process.

[0109] Remove the binding relationship between the target unit and the currently bound heterogeneous computing resources, delete the resource identifier of the heterogeneous computing resources from the unit-resource mapping table, and write the resource identifier into the resource idle tag set to mark it as allocable;

[0110] In simulation systems, temporarily added combat units and heterogeneous computing resources have unit-resource mapping relationships, typically managed through hash tables or associative array structures. To enable adaptive adjustment of binding relationships, the system must immediately unbind a target unit from the current heterogeneous computing resource upon identifying it as being in a rebinding state. This unbinding operation involves searching the unit-resource mapping table, using the unit identifier as an index to extract its corresponding resource identifier, and then deleting the mapping relationship, ensuring the unit is logically unbound from the original resource. Subsequently, the unbound resource identifier is written to a dedicated resource idle flag set, which dynamically maintains a pool of all available idle resources. In practical applications, for example, if target unit U123 is currently bound to GPU-7, when it is determined that U123 needs to be rebound, the system immediately removes GPU-7 from the mapping table and adds it to the resource idle set, indicating that the resource is available for other units to re-match.

[0111] The target unit refers to a temporarily added combat unit that the system currently determines requires adjustment of its binding relationship; it is dynamic and its behavior is complex. Heterogeneous computing resources refer to various computing nodes that provide computing power support in the simulation system, potentially including CPUs, GPUs, FPGAs, etc. The unit-resource mapping table is a data structure used to maintain the binding relationship between units and resources, typically stored in key-value pair format. The resource idle marker set is a set-type structure used by the system to manage the status of allocable resources during operation. Whenever a resource is released, its marker is added to the set, marking it as re-bindingable. Through this unbinding and resource marking mechanism, the system can maintain the continuity and efficiency of resource management, prevent resource leakage and redundant occupation, and ensure the accuracy and real-time nature of subsequent resource allocation.

[0112] Search the resource idle marker set for heterogeneous computing resources that meet the resource adaptation conditions. For each candidate resource, calculate the adaptation degree based on its response capability, historical load level and the short-term behavior change rate index of the target unit. Select the heterogeneous computing resource with the highest adaptation degree as the new binding target, update the unit-resource mapping table, and complete the adaptive adjustment operation of the binding relationship between the temporarily added combat unit and the heterogeneous computing resource.

[0113] During the dynamic scheduling of simulation tasks, to achieve adaptive binding adjustments between temporarily added combat units and heterogeneous computing resources, the system needs to screen and evaluate currently available computing resources based on a resource idle marker set. This is achieved by first retrieving all unoccupied heterogeneous computing resources from the resource idle marker set and then calculating the fit degree for each candidate resource. The fit degree calculation uses the resource's responsiveness parameters (such as execution latency and parallelism metrics), historical load levels (such as CPU or GPU utilization over the past N frames), and the target unit's short-term behavior change rate as input variables to construct a weighted matching function, outputting a fit score that reflects the degree of matching between the unit's dynamic behavior and the resource's responsiveness. For example, if the target unit U456 has a behavior change rate of 0.87 in the recent sliding window, and GPU-3 has high responsiveness and low historical load levels among idle resources, its fit score might be 0.93. If this is the highest among all candidates, GPU-3 is selected as the new binding target. The system then updates the binding relationship between U456 and GPU-3 in the unit-resource mapping table to achieve closed-loop execution of the scheduling logic.

[0114] The resource idle tag set is a dataset used to maintain the identifiers of all currently allocable heterogeneous computing resources. It typically uses a hash set to support fast lookup operations. Resource suitability is a set of evaluation parameters used to measure whether a resource can handle the target unit's computing tasks, including dimensions such as structural performance, response latency, and current load status. Response capability measures the timeliness of a resource's processing of external task requests, historical load levels reflect resource usage within past time windows, and short-term behavior change rate indicators are a quantitative expression of the intensity of behavioral fluctuations of the target unit in a recent period. The suitability calculation function is a scoring mechanism based on a comprehensive evaluation of multiple input factors, often implemented using a weighted linear model, fuzzy inference system, or machine learning model. The unit-resource mapping table is responsible for recording and updating the final binding relationships and is an indispensable core component of the resource scheduling system. Through this process, the system can optimize the matching of computing power to dynamically behaving units while maintaining real-time performance and adaptability.

[0115] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions according to the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means (e.g., infrared, wireless, microwave, etc.). A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0116] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0117] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0118] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0119] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0120] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0121] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A tactical simulation deduction method based on heterogeneous computing power cooperation, characterized in that, Specifically comprising the following steps: S1, by constructing a simulation index structure for distinguishing initialized units from non-initialized units, identifying newly added combat units during simulation running, and marking the newly added combat units as temporarily added combat units based on the identification result; S2, based on the marked temporarily added combat units, collecting state change data and behavior instruction sequences in a plurality of consecutive simulation frames, constructing a behavior event sequence matrix, and calculating the behavior change amount between each frame using a sliding window weighted summation algorithm to establish a short-term behavior change speed index on the behavior event sequence matrix; S3, inputting the short-term behavior change speed index into a resource adaptation evaluation structure, establishing a multivariate mapping relationship between behavior change speed and the response capability of the bound heterogeneous computing resources, extracting a resource response deviation value, and generating a computing power adjustment candidate marker; S4, according to the computing power adjustment candidate marker, constructing a dynamic judgment logic path graph by combining the computing power usage records and behavior change trends of the temporarily added combat units in the previous several simulation frames, and outputting a scheduling decision value of whether the heterogeneous computing power binding relationship needs to be adjusted through conditional judgment of the graph structure; S5, according to the scheduling decision value, performing adaptive adjustment operation on the binding relationship between the temporarily added combat units and the heterogeneous computing resources.

2. The tactical simulation deduction method based on heterogeneous computing power cooperation according to claim 1, characterized in that, S1 specifically comprises: By constructing a simulation index structure for distinguishing initialized units from non-initialized units, generating a key-value pair composed of a unit identifier and a simulation start frame number for each initialized unit during simulation initialization, and writing all key-value pairs into the simulation index structure to form a baseline index set for unit identification during subsequent simulation running period; During simulation running, the unit identifier and the current simulation frame number of the combat unit entering the simulation process are extracted in real time to generate a comparison key-value pair, and the comparison key-value pair is matched with the baseline index set. When the matching result does not exist and the current simulation frame number is greater than the frame number value in any baseline key-value, it is identified as a newly added combat unit; Based on the identification result, the unit identifier and the current simulation frame number of the identified newly added combat unit are written into a new unit record set, and a temporary addition marker is added to the unit identifier for subsequent behavior change speed processing of the temporarily added combat unit.

3. The tactical simulation deduction method based on heterogeneous computing power cooperation according to claim 1, characterized in that, S2 specifically comprises the following steps: S201, based on the marked temporarily added combat units, starting from the simulation frame where the combat unit first appears during simulation running, continuously collecting state change data and behavior instruction sequences in a plurality of simulation frames, and forming a time series set in the order of frame number with the collection results, and archiving storage by unit identifier for subsequent matrix construction; S202, uniformly format coding each frame corresponding state change data and behavior instruction sequence, converting into behavior event vectors of the same dimension, and filling all behavior event vectors into a two-dimensional structure with unit identifier as row index and simulation frame number as column index to construct a behavior event sequence matrix; S203, setting a sliding window structure on the behavior event sequence matrix, using a weighted summation algorithm to calculate the behavior change amount between adjacent frames for each window range, and performing time normalization on the change amount value to form a short-term behavior change speed index, which is used to represent the behavior dynamic intensity of the unit in the window period.

4. The tactical simulation deduction method based on heterogeneous computing power cooperation according to claim 3, characterized in that, S203 specifically includes the following steps: A sliding window structure is constructed on the behavior event sequence matrix at a fixed frame interval, and the width of the sliding window is equal to the number of consecutive simulation frames. The window slides linearly on the behavior event sequence matrix at a frame step, and in each sliding, the behavior event vector set corresponding to all simulation frames in the current window range is extracted as the input basis for behavior change analysis; For each sliding window, the vector difference operation is performed on the behavior event vectors between adjacent frames in the behavior event vector set, and a weighted summation algorithm is performed based on the preset behavior dimension weight configuration, so as to calculate the behavior change amount between each frame in the window range, and generate a behavior change amount sequence as an intermediate output; Each item in the behavior change amount sequence is divided by the corresponding inter-frame time interval to obtain the behavior change intensity value per unit time, and the behavior change intensity value sequence is processed by window mean to generate a short-term behavior change speed index, which is used to represent the behavior dynamic intensity change trend of the temporarily added combat unit in the current sliding window period.

5. The tactical simulation deduction method based on heterogeneous computing power cooperation according to claim 1, characterized in that, S3 specifically includes the following steps: The short-term behavior change speed index is paired with the heterogeneous computing resource identifier bound to the temporarily added combat unit to construct a unit behavior-computing resource input structure, which is input into the resource adaptation evaluation structure. The unit identifier is used as the primary key index to ensure the independence and continuity of the input data of each unit in the evaluation structure; In the resource adaptation evaluation structure, a multivariate mapping construction function is called to uniformly and normalize the short-term behavior change speed index and the response capability parameters of the heterogeneous computing resources, generate a multivariate mapping relationship model reflecting the correlation between the actual load of the resource and the behavior rate of the unit, and calculate the difference value between the current behavior change speed of the unit and the computing response capability based on the multivariate mapping relationship model as the resource response deviation value; The resource response deviation value is compared with the preset deviation threshold interval. When the resource response deviation value exceeds the upper limit of the preset deviation threshold interval, the current unit is marked as a computing adjustment candidate unit, and a computing adjustment candidate marker containing the unit identifier and the deviation value is generated for subsequent judgment and control operation of the resource binding relationship.

6. The tactical simulation deduction method based on heterogeneous computing power cooperation according to claim 1, characterized in that, S4 specifically includes the following steps: S401, according to the unit identifier in the computing adjustment candidate marker, extract the computing power usage record and behavior change trend of the corresponding temporarily added combat unit in the previous several simulation frames, align the two time series according to the frame number, and construct a joint feature array with the frame number as the primary index and the behavior rate and computing power occupation as the joint fields, which is used as the basic input for subsequent path graph construction. S402, constructing a dynamic judgment logic path graph based on the joint feature array, taking frame numbers as path nodes in the graph structure, taking the relationship mode between the behavior change trend and the computing power usage record as the edge weight attribute, connecting the nodes in turn to form a graph path structure, and attaching a judgment condition label to each edge in the path structure to express the associated conditions of the behavior surge and the mismatch of the computing power load; S403, traversing the judgment condition labels in the dynamic judgment logic path graph, identifying path segments that continuously satisfy the edge weight condition anomaly, and performing scoring calculation based on path length, continuous trigger times and edge weight intensity, and when the score value exceeds the adjustment judgment threshold, generating a scheduling decision value for outputting the judgment conclusion of whether the heterogeneous computing power binding relationship needs to be adjusted.

7. The tactical simulation deduction method based on heterogeneous computing power cooperation according to claim 6, characterized in that, S403 is specifically: In the dynamic judgment logic path graph, all path nodes are traversed in time sequence, the judgment condition labels on each path are extracted, and for each edge judgment condition label and edge weight parameter, edge elements that satisfy the abnormal relationship threshold are selected, and their corresponding frame numbers are recorded as potential abnormal marker points; Based on the marked potential abnormal marker points, identify the continuously connected abnormal edge segments in the path graph, calculate the path length, abnormal condition continuous trigger times and corresponding edge weight intensity of each abnormal path segment, and construct a scoring function model through a weighted combination based on logistic regression to generate a path score value reflecting the resource demand pressure level; Compare the path score value with the adjustment judgment threshold, if the path score value is higher than the adjustment judgment threshold, generate a scheduling decision value for the temporarily added combat unit associated with the corresponding path segment, output the judgment conclusion that the heterogeneous computing power binding relationship of the unit needs to be adjusted, and complete the trigger judgment of scheduling state transition.

8. The tactical simulation deduction method based on heterogeneous computing power cooperation according to claim 1, characterized in that, S5 is specifically: According to the scheduling decision value, judge whether the temporarily added combat unit is in the rebinding state, when the scheduling decision value indicates the rebinding state, select the temporarily added combat unit currently in the rebinding state as the target unit, and prepare to perform the adaptive adjustment operation of the binding relationship; Release the binding relationship between the target unit and the currently bound heterogeneous computing resource, delete the resource identifier of the heterogeneous computing resource from the unit-resource mapping table, and write the resource identifier into the resource idle marker set and mark it as allocable state; In the resource idle marker set, search for a heterogeneous computing resource that meets the resource adaptation condition, calculate the adaptation degree of each candidate resource according to its response ability, historical load level and short-time behavior change speed index of the target unit, select the heterogeneous computing resource with the highest adaptation degree as the new binding target, update the unit-resource mapping table, and complete the adaptive adjustment operation of the binding relationship between the temporarily added combat unit and the heterogeneous computing resource.

Images