Feasibility Evaluation Method for Action Plan

Abstract

The present invention discloses a method for assessing the feasibility of action plans, specifically relating to the field of action plan assessment in dynamic confrontation scenarios. It is used to address the problem that the predictability of intelligent agent behavior patterns increases with the extension of mission time, leading to a decrease in the feasibility of action plans. By quantifying behavioral regularity and introducing a perturbation strategy based on terrain and confrontation factors, the method ensures that action adjustments are highly adapted to the environment and maintains the effectiveness of the plan. The application of multiple rounds of deduction and evolutionary confrontation agents in a virtual sandbox enables the early identification and correction of potential vulnerabilities, thereby improving the robustness of the plan. The modular reorganization of tactical actions into a set of new plans enhances the unpredictability of behavior and reduces the possibility of the opponent's reverse deduction of strategies. The real-time monitoring and plan switching mechanism form a closed-loop feedback loop, enabling the intelligent agent to dynamically adapt to changes in the opponent's strategy and extend the feasibility lifecycle of the action plan.

Description

Technical Field

[0001] The present invention relates to the field of action plan evaluation in dynamic confrontation scenarios, and more specifically, to a method for evaluating the feasibility of an action plan. Background Art

[0002] During long-term missions in dynamic adversarial scenarios, intelligent agents must continuously adjust their action strategies based on environmental changes to maintain their effectiveness. As mission duration increases, the agents gradually develop regular behavioral patterns through repeated tactical maneuvers, with predictable characteristics such as path selection and response tempo. This pattern solidification allows the adversary to reverse-engineer the action logic through behavioral analysis, resulting in a gradual decline in the effectiveness of the original plan as the mission progresses, creating a systemic risk of diminishing feasibility over time.

[0003] The current action plan evaluation system is flawed by its lack of a continuous tracking and early warning mechanism for the dynamic evolution of behavioral patterns. When an agent's behavioral characteristics solidify due to long-term tasks, it is impossible to quantitatively assess the predictability risk of its behavior patterns, nor can it trigger targeted policy optimization adjustments. This results in a rigid execution state for action plans that should be dynamically evolving. This disconnect between the evaluation mechanism and behavioral evolution results in the agent continually exposing strategic vulnerabilities during confrontations without being able to autonomously correct them, ultimately irreversibly degrading the overall feasibility of the action plan.

[0004] In order to solve the above problems, a technical solution is now provided. Summary of the Invention

[0005] To overcome the aforementioned shortcomings of the prior art, embodiments of the present invention provide a method for assessing the feasibility of action plans. By quantifying behavioral regularities and introducing perturbation strategies based on terrain and adversarial factors, this method ensures that action adjustments are highly adapted to the environment and maintains the effectiveness of the plan. The application of multiple rounds of deduction and evolutionary adversarial agents in a virtual sandbox enables the early identification and correction of potential vulnerabilities, improving the robustness of the plan. The modular reorganization of tactical actions into a set of new plans enhances the unpredictability of behavior and reduces the possibility of the opponent inferring the strategy in reverse. A closed-loop feedback loop is formed by real-time monitoring and plan switching mechanisms, enabling the agent to dynamically adapt to changes in the opponent's strategy, extending the feasibility lifecycle of the action plan, and thus addressing the issues raised in the aforementioned background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] S1: Collect the agent's path coordinate sequence, task response time, and tactical combination records, and calculate the path selection frequency, time fluctuation, and tactical repetition rate to generate a behavior pattern predictability score;

[0008] S2: Based on the analysis of the terrain morphology oscillation characteristics and the discrete characteristics of the confrontation rhythm, a disturbance intensity adjustment factor is obtained. When the behavior pattern predictability score exceeds the dynamic threshold, the path deviation amplitude and delay time window density are dynamically controlled according to the disturbance intensity adjustment factor to generate a path deviation direction set and disturbance delay time window adapted to the terrain obstacle distribution;

[0009] S3: Using the path offset direction set and perturbation delay time window, the terrain model and adversary strategy library are loaded into the virtual adversarial sandbox to train the adversarial agent with evolutionary capabilities to perform multiple rounds of deduction.

[0010] S4: Extract high-frequency interception position coordinates and tactical sequences during the simulation process, reverse locate exposed path nodes in the decision logic, and generate a set of alternative paths containing modified tactical action sequences;

[0011] S5: Split the corrected tactical action sequence into independent modules according to timestamps, and reorganize it into a new set of discretized behavioral features based on task dependencies;

[0012] S6: Monitor the opponent's strategy recognition rate of the current plan in real time. When the recognition rate exceeds the warning threshold, switch to a new plan that is compatible with the terrain and reset the behavior pattern predictability score.

[0013] In a preferred embodiment, step S1 includes the following contents:

[0014] The position coordinates of the agent during task execution are recorded to form a path coordinate sequence, the time interval from the agent receiving task instructions to the start of execution is recorded to form a task response time sequence, and the tactical action sequence adopted by the agent in task execution is recorded to form a tactical combination record; the task area is divided into critical path segments, and the proportion of the frequency of the agent selecting each critical path segment to the total number of path selections is calculated to obtain the path selection frequency; the task response time concentration is obtained by calculating the ratio of the cumulative change amplitude of adjacent time intervals in the task response time sequence to the difference between the maximum and minimum response times; the tactical repetition is obtained by calculating the proportion of repeated tactical actions in the tactical action sequence to the total number of tactical actions.

[0015] In a preferred embodiment, step S1 further includes the following:

[0016] The path selection frequency is non-dimensionalized to obtain the relative selection frequency; the task response time concentration is non-dimensionalized so that its value is between 0 and 1; the average relative selection frequency of all critical path segments is calculated, and it is added to the non-dimensionalized task response time concentration and tactical repetition, and the added result is mapped to the behavior pattern predictability score.

[0017] In a preferred embodiment, step S2 includes the following:

[0018] Obtain terrain elevation data for the mission area and calculate the terrain morphological oscillation index through frequency domain analysis, where the terrain morphological oscillation index is the energy ratio of the terrain elevation data within a preset spatial frequency range; obtain the opponent's action frequency time series and calculate its coefficient of variation as the confrontation rhythm discrete index; generate a disturbance intensity adjustment factor by weighted summation of the terrain morphological oscillation index and the confrontation rhythm discrete index; set a dynamic threshold, which increases with the increase of the average value of the opponent's action frequency time series; and initiate a disturbance strategy when the behavior pattern predictability score exceeds the dynamic threshold.

[0019] In a preferred embodiment, step S2 further includes the following:

[0020] Calculate the path deviation amplitude, which is proportional to the disturbance intensity adjustment factor and the local obstacle density; calculate the delay time window density, which is proportional to the disturbance intensity adjustment factor and the ratio of the task execution time to the average response time; generate a path deviation direction set, and randomly select directions proportional to the disturbance intensity adjustment factor from the passable directions around the current position; generate disturbance delay time windows, the duration of which is proportional to the inverse of the delay time window density and is evenly distributed on the task timeline.

[0021] In a preferred embodiment, step S3 includes the following contents:

[0022] A three-dimensional terrain grid is constructed by importing terrain elevation data of the mission area, and an opponent strategy model is constructed by importing historical action data of the opponent. The strategy set of the adversarial agent is initialized as the original action plan of the intelligent agent, and a swarm optimization method is used to drive the evolution of the adversarial agent strategy. Based on the fitness function, high-fitness agents are selected for reproduction, parent agent strategy fragments are exchanged with a preset probability for combination, and the tactical actions in the agent strategy are randomly modified with a preset probability for mutation. When applying path offset, the offset direction is randomly selected from the path offset direction set, and the offset amplitude is proportional to the ratio of the total mission path length. When applying delay perturbation, the delay time is randomly selected from the perturbation delay time window, and the delay time is proportional to the ratio of the total mission duration. Multiple rounds of deduction are performed, and the path trajectory, tactical action sequence, and opponent interception events in each round are recorded. The fitness function is composed of the product of the task completion rate and behavioral safety. The task completion rate is the proportion of rounds in which the task is successfully completed, and the behavioral safety is the proportion of rounds in which no interception is achieved. High-frequency interception position coordinates and tactical sequences are extracted from the deduction results.

[0023] In a preferred embodiment, step S4 includes the following contents:

[0024] A density-based clustering method is used to perform cluster analysis on the coordinates of high-frequency interception positions. The interception positions are classified into clusters by setting the neighborhood radius and the minimum number of samples to identify high-incidence areas of interception. The centroid position of each high-incidence area is then calculated, and the path point with the smallest distance to the centroid position is selected from the original path of the intelligent agent and marked as an exposed path node. For each exposed path node, a tactical action subsequence within a preset time window before the interception event occurs is extracted. The opponent strategy model is used to adjust the tactical actions near the exposed path node through an iterative optimization method to reduce the interception probability and generate a modified tactical action sequence. Based on the modified tactical action sequence, a new path segment is generated to replace the exposed path node in the original path. The curve smoothing technique is used to process the connection between the new path segment and the original path to generate a set of alternative paths.

[0025] In a preferred embodiment, step S5 includes the following contents:

[0026] The modified tactical action sequence is divided into independent modules according to timestamps, and each independent module is generated by setting a module duration threshold and accumulating the execution time of the tactical actions; a directed graph is used to model task dependencies, where nodes represent the independent modules obtained by division, and directed edges represent the execution order constraints between independent modules; a module execution sequence that conforms to the task dependency is generated by topological sorting of the directed graph, and a new set of schemes with discretized behavioral characteristics is generated by randomly inserting or replacing independent modules; behavioral characteristic discreteness is introduced as an evaluation indicator, specifically by calculating the ratio of the average behavioral characteristic distance between each new scheme in the new scheme set and the alternative path to the maximum possible distance, and screening out schemes with behavioral characteristic discreteness higher than a preset threshold as new schemes.

[0027] In a preferred embodiment, step S6 includes the following contents:

[0028] The strategy recognition rate is calculated based on the opponent's interception attempt frequency and tactical adjustment speed. The specific operation is to divide the product of the interception attempt frequency and the tactical adjustment speed by the product of the historical maximum value to obtain a ratio. The warning threshold is dynamically adjusted based on the terrain morphology oscillation index and the confrontation rhythm discrete index. The basic threshold is weighted and summed with the normalized values of terrain complexity and confrontation intensity.

[0029] In a preferred embodiment, step S6 further includes the following:

[0030] Schemes compatible with the current terrain are screened from the new scheme set. The difference between the terrain feature vector and the scheme requirement vector is calculated and divided by the sum of their norms to obtain the terrain compatibility index. Schemes with terrain compatibility index higher than the preset threshold are screened out. When the strategy recognition rate exceeds the warning threshold, a scheme is randomly selected from the screened subset of compatible schemes for switching, and the behavior pattern predictability score is reset to the initial value.

[0031] The technical effects and advantages of the action plan feasibility assessment method of the present invention are as follows:

[0032] The present invention provides a method for dynamically evaluating and optimizing the action plans of intelligent agents, which effectively addresses the challenge of increased predictability of behavioral patterns in long-term confrontations. By quantifying behavioral regularities and introducing perturbation strategies for terrain and confrontation factors, it ensures that action adjustments are highly adapted to the environment and maintains the effectiveness of the plan. The application of multiple rounds of deductions and evolutionary confrontation agents in the virtual sandbox enables early identification and correction of potential vulnerabilities, improving the robustness of the plan. The modular reorganization of tactical actions into a set of new plans enhances the unpredictability of behavior and reduces the possibility of the opponent reversely deducing the strategy. The real-time monitoring and plan switching mechanism form a closed-loop feedback, which enables the intelligent agent to dynamically adapt to changes in the opponent's strategy and significantly extends the feasibility life cycle of the action plan. Compared with the traditional static evaluation system, the present invention overcomes the defect of insufficient tracking of behavioral evolution and provides a more adaptable and sustainable solution for dynamic confrontation scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Schematic diagram of the process of the feasibility assessment method of the action plan of the present invention. DETAILED DESCRIPTION

[0034] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0035] Example 1: Figure 1 The feasibility assessment method of the action plan of the present invention is provided, including:

[0036] S1: Collect the agent's path coordinate sequence, task response time, and tactical combination records, and calculate the path selection frequency, time fluctuation, and tactical repetition rate to generate a behavior pattern predictability score;

[0037] S2: Calculate the disturbance intensity adjustment factor based on the terrain morphology oscillation index and the adversarial rhythm dispersion index. When the behavior pattern predictability score exceeds the dynamic threshold, dynamically control the path deviation amplitude and delay time window density according to the disturbance intensity adjustment factor to generate a path deviation direction set and disturbance delay time window that adapts to the terrain obstacle distribution.

[0038] S3: Using the path offset direction set and perturbation delay time window, the terrain model and adversary strategy library are loaded into the virtual adversarial sandbox to train the adversarial agent with evolutionary capabilities to perform multiple rounds of deduction.

[0039] S4: Extract high-frequency interception position coordinates and tactical sequences during the simulation process, reverse locate exposed path nodes in the decision logic, and generate a set of alternative paths containing modified tactical action sequences;

[0040] S5: Split the corrected tactical action sequence into independent modules according to timestamps, and reorganize it into a new set of discretized behavioral features based on task dependencies;

[0041] S6: Monitor the opponent's strategy recognition rate of the current plan in real time. When the recognition rate exceeds the warning threshold, switch to a new plan that is compatible with the terrain and reset the behavior pattern predictability score.

[0042] In dynamic confrontation scenarios, when an agent performs a long-term mission, its behavior pattern gradually exposes predictability risks due to regularities in path selection, response time, and tactical combination. This predictability provides an opportunity for adversaries to reverse engineer the agent's action logic, causing the feasibility of the action plan to decrease over time. To address this issue, step S1 aims to quantify the predictability of the agent's behavior pattern by collecting and analyzing its behavioral data. This provides basic data support for optimizing the perturbation strategy based on terrain and the rhythm of the confrontation in subsequent steps S2 to S6. This step focuses on processing the agent's path coordinate sequence, task response time, and tactical combination records to generate a behavioral pattern predictability score, which serves as input for subsequent perturbation intensity adjustment and path optimization.

[0043] Step S1 includes the following contents:

[0044] S1.1, Data Collection:

[0045] During the data collection phase, the agent's position coordinates during task execution are first recorded, forming a sequence of coordinate points corresponding to multiple timestamps, called a path coordinate sequence. This path coordinate sequence reflects the agent's movement trajectory within the task area. Secondly, the time interval between the agent receiving the task instruction and starting the task is recorded, forming a sequence of response times corresponding to multiple task instructions, called a task response time sequence. This task response time sequence reflects the agent's reaction speed to the task instruction. Finally, the sequence of tactical actions taken by the agent during task execution is recorded, forming a sequence of multiple tactical actions, called a tactical combination record. This tactical combination record reflects the agent's behavioral choices during the confrontation. By comprehensively capturing the agent's movement path, reaction speed, and tactical behavior, it provides a rich data foundation for subsequent analysis, ensuring that the evaluation of behavioral patterns is comprehensive and representative.

[0046] S1.2, Statistical Analysis:

[0047] During the statistical analysis phase, the frequency of path selection is first analyzed. Specifically, the task area is divided into several critical path segments, which are possible connections from the task's starting point to the end point. The selection frequency of each path segment is then calculated as the proportion of the agent's selections of each path segment during the task. This frequency reflects the agent's preference for a particular path. Secondly, the concentration of task response times is analyzed. A concentration index is introduced to measure the degree of concentration in the task response time series. This index calculates the ratio of the cumulative change in adjacent time intervals in the response time series to the difference between the maximum and minimum response times. Lower concentration index values indicate more concentrated response times and more regular agent reaction times. Finally, tactical repetition is analyzed by calculating the proportion of repeated tactical actions in the tactical action sequence. Specifically, this method counts whether each tactical action in the sequence has appeared in the previous action and calculates the proportion of repeated actions in the total number of actions. This proportion reflects the repetitiveness of the agent's tactical choices. The frequency of path selection reveals the degree of dependence of the intelligent agent on a specific path, the concentration of task response time more accurately reflects the regularity of reaction time through indicators, and the tactical repetition directly quantifies the repeatability of behavioral selection.

[0048] S1.3, dimensionless processing:

[0049] During the dimensionless processing phase, the aforementioned statistical indicators are processed to ensure effective mathematical correlation and comparison between different types of data metrics. First, the path selection frequency is dimensionless. This is done by dividing the selection frequency of each path segment by the sum of the selection frequencies of all path segments to obtain the relative selection frequency, so that the sum of the relative selection frequencies of all path segments is a fixed value. Second, the task response time concentration is dimensionless. This is done by converting the concentration index to a value between the minimum and maximum values. The ratio of the concentration index to the maximum possible concentration index is calculated and then subtracted from a fixed value, so that a larger converted index value indicates a more concentrated response time. Finally, the tactical repetition is processed. Since tactical repetition is already a proportion of the total, no additional conversion is required and its original value is used directly. By converting metrics of different dimensions to dimensionless form, mathematical comparability is ensured, facilitating subsequent comprehensive evaluation. Dimensionless processing enhances the physical interpretability of the metrics and avoids interference from dimensional differences in the calculation results.

[0050] S1.4, Calculation of Behavior Pattern Predictability Score:

[0051] During the behavior pattern predictability score calculation phase, the aforementioned dimensionless indicators are combined to calculate a score reflecting the predictability of the agent's behavior pattern. The specific method first calculates the average of the relative selection frequencies of all path segments. This average is then added to the dimensionless task response time concentration and tactical repetition to obtain an intermediate result. Finally, through a nonlinear transformation, this intermediate result is mapped to a behavior pattern predictability score between a minimum and maximum value. The larger the behavior pattern predictability score, the more predictable the agent's behavior pattern. The nonlinear transformation can capture the interactions and complex relationships between indicators and is more adaptable to behavioral changes in dynamic confrontation scenarios than the traditional direct addition method. This improves the accuracy and robustness of the behavior pattern predictability score, ensuring that the evaluation results truly reflect the predictability risk of the behavior pattern.

[0052] The behavioral pattern predictability score serves as input for subsequent steps to determine whether the agent's behavioral strategy needs to be adjusted. Furthermore, the path coordinate sequence and path selection frequency provide the basis for generating data on the direction of subsequent path adjustments, while the task response time series and concentration index provide the basis for determining the density of the behavioral adjustment timeframe.

[0053] Step S1 collects the agent's path coordinate sequence, task response time sequence, and tactical combination records, analyzes path selection frequency, task response time concentration, and tactical repetition, and calculates a behavioral pattern predictability score after dimensionless processing. This score quantifies the predictability risk of the agent's behavior pattern and provides key input for adjusting the perturbation intensity based on the terrain morphology oscillation index and the adversarial rhythm dispersion index in step S2. Furthermore, the path and time-related data lay the foundation for the subsequent path offset and delay window design.

[0054] Step S2 calculates a perturbation intensity adjustment factor based on terrain characteristics and the adversarial rhythm. When the behavioral pattern predictability score exceeds a preset threshold, it dynamically generates a set of path deviation directions and a perturbation delay window adapted to the terrain obstacle distribution, thereby disrupting behavioral regularity and reducing predictability risk. The output of this step, the path deviation direction set and perturbation delay window, provides key input for training in the virtual adversarial sandbox in step S3.

[0055] Step S2 includes the following contents:

[0056] S2.1, Calculation of topographic morphological oscillation index:

[0057] To calculate the Terrain Morphological Oscillation Index (TMOI), we first obtain terrain elevation data for the mission area. This data uses geographic coordinates as independent variables and records the height of each location. The TMOI measures the undulation of the terrain by performing frequency domain analysis on the elevation data and calculating the energy fraction within a preset spatial frequency range. This fraction is calculated by dividing the squared sum of the spectral energy of the elevation data within a specific frequency interval by the total spectral energy. Spectral analysis effectively captures the dynamic characteristics of terrain, particularly in adversarial scenarios, where terrain undulation has a significant impact on the agent's path selection and behavior patterns. This more accurately reflects the impact of terrain on the agent's behavior.

[0058] S2.2, Calculation of the Rhythm Discrepancy Index

[0059] The terrain morphological oscillation features include the Anti-Rhythm Dispersion Index. To calculate this index, we first obtain a time series of the opponent's action frequencies, which records the frequency of the opponent's actions at different points in time. The Anti-Rhythm Dispersion Index measures the volatility of the opposing rhythm by calculating the coefficient of variation (CVO) of the action frequency time series, which is the ratio of the standard deviation of the action frequency to the mean. This technical feature was chosen because the CVO highlights the relative volatility of the opposing rhythm.

[0060] S2.3, Calculation of the disturbance intensity adjustment factor

[0061] The discrete features of the anti-rhythm include a perturbation intensity adjustment factor. This factor is calculated by combining the terrain morphological oscillation index and the anti-rhythm discrete index through a nonlinear fusion approach. The two indices are weighted and summed, and the result is then transformed through a smooth nonlinear function to obtain the perturbation intensity adjustment factor. This nonlinear fusion captures the complex interplay between terrain and anti-rhythm, ensuring that the perturbation intensity is adjusted neither excessively nor insufficiently.

[0062] S2.4, Dynamic Threshold and Disturbance Trigger:

[0063] When setting dynamic thresholds, they adjust based on the intensity of the opponent's actions. This is achieved by adding a modifier proportional to the base threshold and the average frequency of the opponent's actions. The intensity of the opponent's actions directly impacts the predictability risk of the agent's behavior. Dynamic thresholds automatically adjust trigger conditions based on changes in the adversarial environment. This ensures the timely and effective perturbation strategy, avoiding unnecessary perturbations when the adversary's intensity is low while increasing perturbations when the adversary is intense.

[0064] S2.5, Dynamic Control of Path Deviation Amplitude and Delay Time Window Density:

[0065] When dynamically controlling the path deviation amplitude, the deviation amplitude is proportional to the disturbance intensity adjustment factor and the local obstacle density. This is achieved by multiplying the disturbance intensity adjustment factor by the ratio of the local obstacle density to the regional average obstacle density. This is because areas with dense obstacles require smaller deviation amplitudes to prevent the agent from getting stuck in unfavorable terrain due to excessive deviation. This control method allows the path deviation to be highly adapted to the terrain characteristics, enhancing the unpredictability of the agent's behavior while ensuring operational safety.

[0066] When dynamically controlling the density of the delay window, the density is proportional to the perturbation intensity adjustment factor and the ratio of the average task execution time to the average response time. Task execution time and response time reflect the agent's behavioral rhythm, and density control can adjust the frequency of perturbations based on the characteristics of the task. Matching the delay window setting to the task rhythm ensures the tempo of perturbations.

[0067] S2.6, generate the path deviation direction set and disturbance delay time window:

[0068] When generating a set of path deviation directions, the agent first calculates the angular range of possible directions around the current position, excluding directions blocked by obstacles. Then, a number of directions are randomly selected from these possible directions, with the number of directions proportional to the value of the perturbation intensity adjustment factor. This random selection of directions increases the unpredictability of the path while also accounting for terrain obstacles and ensuring the feasibility of the deviation direction. This makes the agent's path selection more flexible and diverse, reducing the risk of being predicted by an opponent.

[0069] When generating perturbation delay windows, their duration is proportional to the inverse of the delay window density and is evenly distributed across the task timeline. These evenly distributed windows cover the entire task cycle, ensuring the temporal comprehensiveness of the perturbation and making it more difficult for adversaries to discern patterns in the agent's behavior over time.

[0070] Step S3 uses the path offset direction set and perturbation delay time window to load the terrain model and opponent strategy library in the virtual confrontation sandbox, and trains the evolutionary adversarial agent to perform multiple rounds of deductions. The purpose is to evaluate and optimize the effectiveness of the perturbation strategy and provide key data such as high-frequency interception position coordinates and tactical sequences for step S4.

[0071] Step S3 includes the following contents:

[0072] S3.1, Construction of Virtual Confrontation Sandbox:

[0073] When building a virtual combat sandbox, the first step is to import terrain elevation data for the mission area. This data records the altitude of each geographic location and constructs a three-dimensional terrain grid based on these elevation values. The grid's horizontal and vertical resolutions are fixed at a pre-set interval. Next, historical adversary action data is imported and used to construct an adversary strategy model. This strategy model employs a network structure, where nodes represent possible adversary tactical options and edges represent the conditional dependencies between these options. By combining terrain elevation data with the adversary strategy model, a realistic combat environment can be dynamically simulated, enhancing the realism and complexity of the simulation. This provides a high-fidelity simulation platform for subsequent training and simulations, ensuring the reliability and credibility of the training results.

[0074] S3.2, Initialization and evolution mechanism of adversarial agents:

[0075] When initializing the adversarial agent, its initial strategy set is set to the agent's original action plan, ensuring that training begins with the agent's actual behavioral patterns. Within the evolutionary mechanism, a swarm optimization-based approach drives the improvement of the adversarial agent's strategy. The agent swarm consists of multiple individuals, each representing a strategy variant. Specific operations include: selecting agents with higher fitness from the swarm for reproduction based on predefined fitness criteria; combining two parent agents by exchanging partial strategy fragments with a preset probability; and randomly modifying a tactical action in the agent's strategy with a preset probability for mutation. This approach simulates the selection and evolutionary processes in nature, enabling the adversarial agent to gradually optimize its strategy within the virtual environment to adapt to terrain changes and the opponent's strategic adjustments.

[0076] S3.3, Application of Path Deviation and Delay Perturbation:

[0077] When applying path offsets, the adversarial agent randomly selects an offset direction from a set of path offset directions during task execution and adjusts its path according to a preset offset magnitude. The offset magnitude is proportional to the ratio of the total task path length to ensure that the offset magnitude matches the task scale and maintains rationality. When applying delay perturbations, after receiving the task instruction, the adversarial agent randomly selects a delay time from the perturbation delay time range and begins task execution only after that time. The delay time is proportional to the ratio of the total task duration to adapt to the task time scale. By associating path offsets and delay perturbations with the spatiotemporal scale of the task, the applicability and effectiveness of the perturbation strategy in different task scenarios can be ensured, making the perturbation strategy more targeted and enhancing the unpredictability of the agent's behavior while maintaining the rationality and controllability of the actions.

[0078] S3.4, Execution and Data Recording of Multiple Rounds of Simulations:

[0079] During multiple rounds of simulation, each lasting a pre-set simulation time, the adversarial agent's path, tactical action sequences, and adversary interception events are recorded. An interception event occurs when the adversary successfully deploys a tactic along the adversarial agent's path, resulting in mission failure or delay. Through multiple rounds of simulation, the dynamic evolution of the adversarial agent's behavior patterns and the adversary's response patterns can be captured, providing rich data support for subsequent analysis. This long-term simulation can reveal temporal trends in behavioral patterns, providing a more comprehensive data basis for optimizing strategies.

[0080] S3.5, design of fitness function:

[0081] When constructing the fitness function, the fitness value is composed of two parts. The first part is the proportion of rounds in which the adversarial agent successfully completes tasks, which is used to measure the task completion rate. The second part is the proportion of rounds in which the adversarial agent is not intercepted by the opponent, which is used to measure behavioral safety. The final fitness value is obtained by multiplying these two parts. By multiplying the task completion rate and behavioral safety, the overall performance of the adversarial agent strategy can be nonlinearly reflected. This allows the fitness function to more accurately assess the overall performance of the adversarial agent in the adversarial environment, thereby guiding the evolutionary process towards more optimal strategies.

[0082] The specific method for determining high-frequency interception location coordinates is to count the location coordinates of enemy interception events during multiple rounds of simulations. These locations are then defined as those where the number of interceptions exceeds a preset threshold. Specifically, during multiple simulations, the path trajectories, tactical action sequences, and enemy interception events are recorded in each round. By statistically analyzing the location data of interception events across all simulation rounds, locations where the number of interceptions reaches or exceeds a predetermined threshold are selected to identify them as high-frequency interception points. This process relies on the statistical results of multiple simulation rounds to ensure objectivity and reliability.

[0083] The high-frequency interception position coordinates and corresponding tactical sequences are extracted from the results of multiple rounds of simulation. This data is directly passed to step S4 for analysis of potential path nodes exposed in the decision logic. Simultaneously, the evolution of the adversarial agent strategy during the simulation is passed to step S5 as input for generating a new set of action plans.

[0084] Step S4 aims to reversely locate the exposed path nodes in the decision logic from the high-frequency interception position coordinates and tactical sequence, and generate a set of alternative paths containing the modified tactical action sequence, providing the basis for the optimized action plan for step S5.

[0085] Step S4 includes the following contents:

[0086] S4.1, Cluster analysis of high-frequency interception position coordinates:

[0087] When performing cluster analysis on the coordinates of high-frequency interception locations, a density-based clustering method is first used to process the location data of interception events. Dense areas in the data are identified by setting a neighborhood radius and a minimum number of samples. Specifically, for each interception location point, the number of other locations within its neighborhood radius is checked. If the number of other locations exceeds the minimum number of samples, these points are grouped into a cluster and expanded to adjacent dense areas. The density-based clustering method can adaptively identify the different shapes and sizes of high-incidence interception areas, more accurately reflecting the complex distribution of terrain characteristics and mission environments, and providing a reliable regional division basis for optimizing exposure paths.

[0088] S4.2, reverse positioning of exposed path nodes:

[0089] To reverse-engineer the exposed path nodes, the center of mass of each high-interception zone is first calculated. The horizontal and vertical coordinates of all interception locations within that zone are averaged to obtain the center of mass coordinates. Next, the distance between each path point and the center of mass is calculated for each path point in the agent's original path. The path point with the smallest distance is selected and marked as an exposed path node. By matching the distance between the center of mass and the path point, the weak points in the agent's path that are most relevant to the high-interception zone can be directly located. This allows for rapid identification of path locations requiring optimization, providing clear improvement targets for tactical adjustments.

[0090] S4.3, Feature extraction of tactical sequences:

[0091] To extract tactical sequence features, for each exposed path node, the tactical action subsequence preceding the interception event is analyzed. Starting from the interception event's time point, the system traces back a pre-set time window, extracting all tactical actions executed by the agent within that time period to form an action subsequence. Since tactical actions preceding an interception event are often the key factor leading to exposure, tracing back within a fixed time window allows for systematic capture of these key behavioral patterns. This facilitates the precise identification of tactical features directly related to the interception event, providing a specific and targeted data foundation for optimization.

[0092] S4.4, correct the generation of tactical action sequences:

[0093] When generating a modified tactical action sequence, the adversary's strategy model is leveraged to employ an optimization method to adjust tactical actions near exposed path nodes to reduce the probability of interception. Specifically, the interception probability is first calculated by calculating the ratio of the number of interceptions near exposed path nodes to the total number of missions executed. Then, the interception probability is gradually reduced by iteratively adjusting the tactical actions until a preset optimization threshold is met. By integrating this optimization with the adversary's strategy model, the modified tactical actions are made more targeted and adaptable to the adversary's interception behavior. This significantly reduces the agent's exposure risk during confrontation, thereby improving the stealth and success rate of the action plan.

[0094] S4.5, Generation of a set of alternative paths:

[0095] When generating a set of alternative paths, new path segments are generated based on the revised tactical action sequence, replacing exposed path nodes in the original path. Path segments are replanned based on the revised action sequence, and curve smoothing techniques are used to smooth the junctions between the replaced paths, maintaining path continuity and smoothness. This path segment replacement and smoothing effectively avoid exposed path nodes while ensuring the overall smoothness of the path, meeting mission execution requirements. This improves path safety while ensuring the agent's operational efficiency and mission feasibility.

[0096] The generated set of alternative paths and the revised tactical action sequence are passed as candidate data to step S5 for generating a new set of action plans.

[0097] Step S4 uses cluster analysis to identify high-incidence interception areas based on the high-frequency interception location coordinates from step S3 and reverse engineer the exposed path nodes in the agent's decision logic. Based on the tactical sequence from step S3, key tactical subsequences are extracted and optimized using the adversary strategy model to generate a revised tactical action sequence. These revised sequences generate a set of alternative paths, providing the basis for an optimized action plan in step S5. This ensures that the new plan effectively avoids exposed nodes, improving both stealth and feasibility.

[0098] Step S5 aims to collect the candidate paths

[0099] The modified tactical action sequence in is split into independent modules according to timestamps and reorganized into a new set of solutions with discretized behavioral features based on task dependencies, providing switchable action plans for step S6.

[0100] Step S5 includes the following contents:

[0101] S5.1, corrected the splitting of tactical action sequences:

[0102] When splitting a sequence of corrected tactical actions, the corrected tactical action sequence in each alternative path is first divided according to its timestamp. The division method is as follows: based on the execution time interval of each tactical action, a fixed module duration threshold is set. Starting from the first action in the sequence, the execution time of each action is accumulated. When the accumulated total time reaches or exceeds the module duration threshold, this part of the action is divided into an independent module. The time accumulation is restarted from the next action, and this process is repeated until the entire sequence is processed. This dynamic division based on time windows ensures that each independent module contains sufficient behavioral information to reflect the tactical intention of the agent, while maintaining the independence of the independent modules, facilitating subsequent reorganization and adjustment. This splitting method enhances the flexibility of the independent modules, enabling them to adapt to the needs of different mission scenarios and laying the foundation for the discretization of behavioral characteristics.

[0103] S5.2, Modeling of Task Dependencies:

[0104] When modeling mission dependencies, a directed graph is used to represent the execution order and conditional constraints between tactical action modules. This approach treats each independent module as a node. If the execution of one independent module requires the completion of another, a directed edge is added between the two independent modules to clearly indicate the execution order. Dependencies are determined based on the logical sequence of tactical actions and the specific requirements of the mission objectives.

[0105] S5.3, Reorganization of discretized behavioral features:

[0106] When reorganizing the discretized behavioral features, the directed graph of task dependencies is first topologically sorted to generate a module execution sequence that conforms to the dependencies and ensures that the order of independent modules meets logical constraints. Based on this, a new execution sequence is generated by randomly inserting or replacing independent modules. Specifically, independent modules compatible with the current task are selected from other alternative paths and their order or combination is randomly adjusted to form a new solution. This reorganization method not only meets the basic requirements of task execution, but also makes the behavioral characteristics of the new solution significantly different from the original solution, thereby improving the solution's stealth and adversarial capabilities.

[0107] S5.4, Evaluation and Screening of New Solution Sets:

[0108] When evaluating and screening new solution sets, behavioral feature dispersion is introduced as an evaluation metric. The average behavioral feature distance between the new solution and each path in the set of alternative paths is calculated, where the behavioral feature distance is defined based on the differences in the execution order and action type of independent modules. The average of all distances is then compared with the theoretical maximum possible distance to obtain the behavioral feature dispersion. The behavioral feature dispersion quantifies the degree of difference between the new solution and the original solution, thereby ensuring that the new solution is sufficiently unique. By selecting solutions with behavioral feature dispersion above a preset threshold as new solutions, the diversity and concealment of the new solution set can be guaranteed, providing high-quality alternative solutions for solution switching.

[0109] The resulting set of new solutions is organized into a structured library of alternatives, which are then fed directly into the subsequent processing steps. Providing a diverse set of alternatives ensures rapid response to adversary strategy changes in dynamic confrontation scenarios, thereby maintaining the effectiveness of the action plan.

[0110] Step S6 includes the following contents:

[0111] S6.1, Real-time monitoring of strategy recognition rate:

[0112] When monitoring the strategy recognition rate in real time, the strategy recognition rate is first calculated using the opponent's action feedback data. The number of interception attempts per unit time by the opponent's tactical adjustment frequency is multiplied. This product is then compared with the product of the maximum interception attempt frequency and tactical adjustment speed in the historical record. The resulting product is the strategy recognition rate. The number of interception attempts reflects the opponent's attention to the agent's behavior, while the tactical adjustment frequency reflects the speed at which the opponent adapts to the agent's strategy. The product of the two comprehensively represents the depth of the opponent's recognition of the agent's action plan, enabling more sensitive capture of subtle changes in the opponent's behavior in dynamic adversarial environments.

[0113] S6.2, Dynamic Adjustment of Alert Thresholds:

[0114] When dynamically adjusting the alert threshold, the threshold is adjusted based on terrain complexity and adversarial intensity. The specific method is to perform a weighted summation of the base threshold with the normalized values of terrain complexity and adversarial intensity to obtain the final alert threshold. Terrain complexity is represented by the Terrain Morphology Oscillation Index, and adversarial intensity is represented by the Adversary Rhythm Discrete Index. Changes in terrain complexity and adversarial intensity directly affect the predictability risk of the agent's behavior. By dynamically adjusting the alert threshold, flexible responses can be implemented based on specific scenarios, ensuring more cautious triggering of action plan switching in highly complex or high-intensity environments. This improves the environmental adaptability of the alert threshold, avoiding the misjudgments or delayed responses that can result from fixed thresholds in different scenarios, thereby enhancing the accuracy and timeliness of decision-making.

[0115] S6.3, Screening of new terrain compatible solutions:

[0116] When screening new solutions for terrain compatibility, the team first extracts terrain features from the current mission area, including slope and obstacle density, and organizes these features into a vector. Simultaneously, for each solution in the new solution set, the terrain requirements for execution are estimated and a corresponding requirement vector is generated. A compatibility index is then calculated by calculating the ratio of the difference between the terrain feature vector and the solution requirement vector to the sum of their norms. By quantifying the degree of match between terrain features and solution requirements, the selected solutions are ensured to have high feasibility in the current terrain environment. This screening method improves the success rate of action plan switching, avoids execution failures due to terrain mismatch, and thus enhances the agent's survivability in complex terrain environments.

[0117] S6.4, Scheme Switching and Behavior Pattern Predictability Score Reset:

[0118] When switching strategies, if the strategy recognition rate exceeds the warning threshold, a new strategy is randomly selected from a subset of new strategies that meet terrain compatibility requirements and replaced with it. Simultaneously, the behavioral pattern predictability score is reset to an initial value to reflect the low predictability of the new strategy. Randomly selecting a new strategy increases the uncertainty of the switch, preventing the adversary from further deciphering the agent's behavioral logic through pattern recognition. Resetting the behavioral pattern predictability score ensures the continuity of the evaluation process and promptly reflects the stealth of the new strategy. This ensures the real-time and effectiveness of the evaluation mechanism while maintaining the stealth of the action plan, providing technical support for the agent's continuous optimization in adversarial environments.

[0119] Step S6 monitors the opponent's strategy recognition rate in real time. When the strategy recognition rate exceeds the dynamically adjusted warning threshold, it selects a terrain-compatible solution from the new solution set in step S5 and switches to it. It also resets the behavior pattern predictability score. This ensures that the agent can promptly respond to the opponent's strategy recognition in dynamic confrontation scenarios, maintaining the concealment and feasibility of the action plan, thereby improving the success rate of mission execution.

[0120] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0121] It should be noted that the system of the present invention can be deployed on the device itself to realize embedded applications, and can also be run on a PC or other terminal with a user interface, thereby meeting a variety of hardware environments and usage requirements.

[0122] The above description is merely illustrative of certain exemplary embodiments of the present invention. It goes without saying that those skilled in the art will be able to modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and description are illustrative in nature and should not be construed as limiting the scope of protection of the claims.

[0123] It should be noted that, in this document, if there are relational terms such as first and second, etc., they are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article or device. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device that includes the element.

[0124] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. The feasibility assessment method of an action plan is characterized by: Including steps: S1: Collect the agent's path coordinate sequence, task response time, and tactical combination records, and calculate the path selection frequency, time fluctuation, and tactical repetition rate to generate a behavior pattern predictability score; S2: Based on the analysis of the terrain morphology oscillation characteristics and the discrete characteristics of the confrontation rhythm, a disturbance intensity adjustment factor is obtained. When the behavior pattern predictability score exceeds the dynamic threshold, the path deviation amplitude and delay time window density are dynamically controlled according to the disturbance intensity adjustment factor to generate a path deviation direction set and disturbance delay time window adapted to the terrain obstacle distribution; S3: Using the path offset direction set and perturbation delay time window, the terrain model and adversary strategy library are loaded into the virtual adversarial sandbox to train the adversarial agent with evolutionary capabilities to perform multiple rounds of deduction. S4: Extract high-frequency interception position coordinates and tactical sequences during the simulation process, reverse locate exposed path nodes in the decision logic, and generate a set of alternative paths containing modified tactical action sequences; S5: Split the corrected tactical action sequence into independent modules according to timestamps, and reorganize it into a new set of discretized behavioral features based on task dependencies; S6: Monitor the opponent's strategy recognition rate of the current plan in real time. When the recognition rate exceeds the warning threshold, switch to a new plan that is compatible with the terrain and reset the behavior pattern predictability score.

2. The action plan feasibility assessment method according to claim 1, characterized in that: Step S1 includes the following contents: The position coordinates of the agent during task execution are recorded to form a path coordinate sequence, the time interval from the agent receiving task instructions to the start of execution is recorded to form a task response time sequence, and the tactical action sequence adopted by the agent in task execution is recorded to form a tactical combination record; the task area is divided into critical path segments, and the proportion of the frequency of the agent selecting each critical path segment to the total number of path selections is calculated to obtain the path selection frequency; the task response time concentration is obtained by calculating the ratio of the cumulative change amplitude of adjacent time intervals in the task response time sequence to the difference between the maximum and minimum response times; the tactical repetition is obtained by calculating the proportion of repeated tactical actions in the tactical action sequence to the total number of tactical actions.

3. The action plan feasibility assessment method according to claim 2, characterized in that: Step S1 also includes the following: The path selection frequency is non-dimensionalized to obtain the relative selection frequency; the task response time concentration is non-dimensionalized so that its value is between 0 and 1; the average relative selection frequency of all critical path segments is calculated, and it is added to the non-dimensionalized task response time concentration and tactical repetition, and the added result is mapped to the behavior pattern predictability score.

4. The action plan feasibility assessment method according to claim 3, characterized in that: Step S2 includes the following contents: Obtain terrain elevation data for the mission area and calculate the terrain morphological oscillation index through frequency domain analysis, where the terrain morphological oscillation index is the energy ratio of the terrain elevation data within a preset spatial frequency range; obtain the opponent's action frequency time series and calculate its coefficient of variation as the confrontation rhythm discrete index; generate a disturbance intensity adjustment factor by weighted summation of the terrain morphological oscillation index and the confrontation rhythm discrete index; set a dynamic threshold, which increases with the increase of the average value of the opponent's action frequency time series; and initiate a disturbance strategy when the behavior pattern predictability score exceeds the dynamic threshold.

5. The action plan feasibility assessment method according to claim 4, characterized in that: Step S2 also includes the following: Calculate the path deviation amplitude, which is proportional to the disturbance intensity adjustment factor and the local obstacle density; Calculate the delay time window density, which is proportional to the disturbance intensity adjustment factor and the ratio of the task execution time to the average response time; generate a path offset direction set, and randomly select directions proportional to the disturbance intensity adjustment factor from the passable directions around the current position; generate disturbance delay time windows, the duration of which is proportional to the inverse of the delay time window density and is evenly distributed on the task timeline.

6. The action plan feasibility assessment method according to claim 5, characterized in that: Step S3 includes the following contents: A three-dimensional terrain grid is constructed by importing terrain elevation data of the mission area, and an opponent strategy model is constructed by importing historical action data of the opponent. The strategy set of the adversarial agent is initialized as the original action plan of the intelligent agent, and a swarm optimization method is used to drive the evolution of the adversarial agent strategy. Based on the fitness function, high-fitness agents are selected for reproduction, parent agent strategy fragments are exchanged with a preset probability for combination, and the tactical actions in the agent strategy are randomly modified with a preset probability for mutation. When applying path offset, the offset direction is randomly selected from the path offset direction set, and the offset amplitude is proportional to the ratio of the total mission path length. When applying delay perturbation, the delay time is randomly selected from the perturbation delay time window, and the delay time is proportional to the ratio of the total mission duration. Multiple rounds of deduction are performed, and the path trajectory, tactical action sequence, and opponent interception events in each round are recorded. The fitness function is composed of the product of the task completion rate and behavioral safety. The task completion rate is the proportion of rounds in which the task is successfully completed, and the behavioral safety is the proportion of rounds in which no interception is achieved. High-frequency interception position coordinates and tactical sequences are extracted from the deduction results.

7. The action plan feasibility assessment method according to claim 6, characterized in that: Step S4 includes the following contents: A density-based clustering method is used to perform cluster analysis on the coordinates of high-frequency interception positions. The interception positions are classified into clusters by setting the neighborhood radius and the minimum number of samples to identify high-incidence areas of interception. The centroid position of each high-incidence area is then calculated, and the path point with the smallest distance to the centroid position is selected from the original path of the intelligent agent and marked as an exposed path node. For each exposed path node, a tactical action subsequence within a preset time window before the interception event occurs is extracted. The opponent strategy model is used to adjust the tactical actions near the exposed path node through an iterative optimization method to reduce the interception probability and generate a modified tactical action sequence. Based on the modified tactical action sequence, a new path segment is generated to replace the exposed path node in the original path. The curve smoothing technique is used to process the connection between the new path segment and the original path to generate a set of alternative paths.

8. The action plan feasibility assessment method according to claim 7, characterized in that: Step S5 includes the following contents: The modified tactical action sequence is divided into independent modules according to timestamps, and each independent module is generated by setting a module duration threshold and accumulating the execution time of the tactical actions; a directed graph is used to model task dependencies, where nodes represent the independent modules obtained by division, and directed edges represent the execution order constraints between independent modules; a module execution sequence that conforms to the task dependency is generated by topological sorting of the directed graph, and a new set of schemes with discretized behavioral characteristics is generated by randomly inserting or replacing independent modules; behavioral characteristic discreteness is introduced as an evaluation indicator, specifically by calculating the ratio of the average behavioral characteristic distance between each new scheme in the new scheme set and the alternative path to the maximum possible distance, and screening out schemes with behavioral characteristic discreteness higher than a preset threshold as new schemes.

9. The action plan feasibility assessment method according to claim 8, characterized in that: Step S6 includes the following contents: The strategy recognition rate is calculated based on the opponent's interception attempt frequency and tactical adjustment speed. The specific operation is to divide the product of the interception attempt frequency and the tactical adjustment speed by the product of the historical maximum value to obtain a ratio. The warning threshold is dynamically adjusted based on the terrain morphology oscillation index and the confrontation rhythm discrete index. The basic threshold is weighted and summed with the normalized values of terrain complexity and confrontation intensity.

10. The action plan feasibility assessment method according to claim 9, characterized in that: Step S6 also includes the following: Schemes compatible with the current terrain are screened from the new scheme set. The difference between the terrain feature vector and the scheme requirement vector is calculated and divided by the sum of their norms to obtain the terrain compatibility index. Schemes with terrain compatibility index higher than the preset threshold are screened out. When the strategy recognition rate exceeds the warning threshold, a scheme is randomly selected from the screened subset of compatible schemes for switching, and the behavior pattern predictability score is reset to the initial value.

Images