Artificial intelligence confrontation technology military training simulation system based on reinforcement learning

Abstract

The invention provides an artificial intelligence confrontation technology military training simulation system based on reinforcement learning, and belongs to the field of training simulation. The problem of poor confrontation simulation training effect is solved; the scene simulation module is used for constructing a differential equation set and setting a response area of a target or a response interval of a laser module; the training simulation module is used for acquiring the positions of trainees, counting, recording and analyzing the shooting interval and the effective score of each trainee in combination with a differential equation set and the response area of the target or the response interval of the laser module, and adjusting a training strategy; the score feedback module is used for recording and feeding back effective scores of the trainees; the target refreshing interval and position are adjusted by analyzing the score and shooting angle of the trainee, and the target moving route is distributed according to the position of the trainee, so that the model effect of confrontation training is improved.

Description

Technical Field

[0001] This invention relates to a military training simulation system based on reinforcement learning and artificial intelligence adversarial technology, and pertains to the field of training simulation. Background Technology

[0002] Existing simulation systems for adversarial training have the following shortcomings: Lack of realistic shooting behavior characteristics: The biggest problem with some laser weapon simulation systems is that they often simplify "shooting" to a geometric straight line, ignoring the many complex physical and physiological factors in real shooting behavior. This idealized model leads to a systematic deviation between training results and actual combat performance, limiting its application depth in modern training systems.

[0003] Blind spots for trainees cannot be automatically identified: In traditional training systems, shooting blind spots (such as posture blind spots, habitual aiming blind spots, and reaction delay blind spots) cannot be identified; most training systems can only record whether a shot is hit, but cannot record key characteristics such as whether the direction is covered, whether the coverage density is insufficient, and whether the blind spot is statistically significant. This makes the blind spot analysis of trainees completely dependent on the instructor's experience and observation; however, the instructor's experience and observation are subjective, fragmented, and unstable, and it is impossible to track and accurately quantify each trainee over a long period of time.

[0004] The training difficulty cannot be dynamically adjusted according to performance: Most laser simulation training modes adopt "fixed difficulty", such as a fixed number of targets, a fixed refresh rate, and a fixed time interval between appearances. This training method ignores the real-time performance and ability fluctuations of trainees, making the trainees' operation tend to be mechanical and unable to produce real skill growth. Summary of the Invention

[0005] In view of the shortcomings of existing technologies, the purpose of this invention is to provide a military training simulation system for artificial intelligence adversarial technology based on reinforcement learning, which aims to solve the problem of poor adversarial simulation training results.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A military training simulation system for artificial intelligence adversarial technology based on reinforcement learning includes: Scene simulation module: used to simulate the flight speed and trajectory of flying objects, and to construct differential equations 1 and 2; based on the activity status of the trainees and their actual positions, to set the response area of ​​the target or the response interval of the laser module; Training simulation module: used to acquire the position of trainees in real time, and combine differential equations 1 and 2 with the response area of ​​the target or the response interval of the laser module to count and record the shooting interval and effective score of each trainee; based on the shooting interval and effective score of all trainees, execute the fixed target training submodule, moving target training submodule and mutual confrontation training submodule in sequence. The training simulation module includes: a fixed target training submodule, a moving target training submodule, and a mutual adversarial training submodule; Fixed Target Training Submodule: Used to acquire and analyze the shooting interval of trainees shooting at ordinary targets, and adjust the refresh interval of interference targets; acquire and analyze the number of shots, shooting interval and hit rate of trainees shooting at interference targets, adjust the shooting interval of dangerous targets, and determine whether the effective scores of all trainees are qualified; if qualified, the moving target training submodule is executed; if unqualified, training continues. Moving Target Training Submodule: This module randomly refreshes ordinary and distracting targets around the trainees and determines whether the effective scores of all trainees are qualified. If they are not qualified, training continues. If they are qualified, dangerous targets are refreshed around the trainees based on their shooting angles and effective scores. The effective scores of all trainees are then tallied and determined again. If they are qualified, the mutual confrontation training submodule is executed. If they are not qualified, training continues. The mutual confrontation training submodule is used to divide dangerous targets into advance targets and support targets, and set the maximum number of advance targets to be deployed and the maximum number of targets to be gathered; adjust the movement paths of advance targets and support targets according to the location of the trainees; count the effective scores of all trainees, and retrain the trainees with unsatisfactory effective scores until they pass the effective score. Performance Feedback Module: Used to statistically analyze and record the effective performance of all trainees, summarize the information of trainees who participated in intensive combat training, and provide feedback.

[0007] Furthermore, the training simulation module includes the following sub-modules: Fixed Target Training Submodule: Used to acquire and analyze the shooting interval of trainees shooting at ordinary targets, and adjust the refresh interval of interference targets; acquire and analyze the number of shots, shooting interval and hit rate of trainees shooting at interference targets, adjust the shooting interval of dangerous targets, and determine whether the effective scores of all trainees are qualified; if qualified, the moving target training submodule is executed; if unqualified, training continues. Moving Target Training Submodule: This module randomly refreshes ordinary and distracting targets around the trainees and determines whether the effective scores of all trainees are qualified. If they are not qualified, training continues. If they are qualified, dangerous targets are refreshed around the trainees based on their shooting angles and effective scores. The effective scores of all trainees are then tallied and determined again. If they are qualified, the mutual confrontation training submodule is executed. If they are not qualified, training continues. The mutual confrontation training submodule is used to divide dangerous targets into advance targets and support targets, and set the maximum number of advance targets to be deployed and the maximum number of targets to be gathered; adjust the movement paths of advance targets and support targets according to the location of the trainees; count the effective scores of all trainees, and conduct confrontation training again for trainees with unsatisfactory effective scores until the effective scores are satisfactory.

[0008] Furthermore, the workflow of the scene simulation module is as follows: Based on the speed v of the launch device （o） Based on the horizontal angle, elevation angle, and the weight of the simulated flying object, construct differential equations 1 and 2; Get the location of the trainees (x) （tr） y （tr） , z （tr） ) and bullseye position (x （ta） y （ta） , z （ta） ), obtain the wind speed (wx, wy, wz) of the training field; and set the target response time interval based on the activity status of the trainees and their actual positions. Scenario 1: The trainees remain stationary; Calculate the actual distance l from the trainee to the target based on the trainee's position and the target's position. （dis） Calculate the vertical distance l from the trainee to the target center. （ver） ; Determine if there is wind in the training area where the trainees are located; If not, the target response area and laser module response interval are not set, and only the laser module operated by the trainee is checked to see if it hits the bullseye. If so, the flight trajectory of the flying object is approximated by a parabolic trajectory, and the response interval of the target and the response interval of the laser module are set. Calculate the ideal flight time t of the flying object （ide） ; Let the horizontal angle at the time of launch be θ. （1） The elevation angle is φ （1） ; Construct equation 1-1: ; The horizontal angle θ can be obtained by solving equation 1-1.（1） and elevation angle φ （1） The target's response region is obtained: ; Calculate the time t for the laser to reach the target. （las） ; The response interval of the laser module is t. （del） : ; Scenario 2: The trainees can move.

[0009] Furthermore, the target response time interval for scenario two is set as follows: Obtain the slowest reflex arc reaction time t of the trainee （ref） ; Determine whether the distance from the trainee to any target center is within the standard distance l （sta） Within; standard distance l （sta） : ; If a trainee's distance to the nearest target's bullseye is within the standard distance l （sta） Within this range, no target response area or laser module response interval is set; only the laser module operated by the trainee is tested to see if it hits the bullseye. If the distance between the trainee and the nearest target bullseye is not within the standard distance... （sta） If the distance is within 1 lb, then calculate the distance lb from the trainee to the bullseye. （tar） ; The response interval of the laser module is ti. （tar） : ; Scenario 2: No target response area is set; Based on differential equations 1 and 2 and the response region of the target or the response interval of the laser module, the shooting interval and effective score of each trainee shooting at a normal target are statistically recorded, and the fixed target training sub-module is executed.

[0010] Furthermore, the workflow of the fixed target training submodule is as follows: Calculate the percentage of trainees who passed and failed, and then calculate the weighted average score (wg). （a） ; The weighted effective score (wg) of all trainees during fixed-target shooting is determined. （a2） Is it qualified? If the results are unsatisfactory, then trainees with unsatisfactory results will be trained. If it passes the test, adjust the refresh interval of the jamming target. The firing interval and hit rate of all trainees were standardized, and then the correlation coefficient between firing interval and hit rate was calculated. Determine if the absolute value of the correlation coefficient is greater than 0.8; If it is greater than that, then a multinomial regression equation is constructed with the shooting interval of all trainees as the independent variable and the hit rate of all trainees as the dependent variable. Calculate the numbers tt that minimize and maximize the polynomial regression equation, respectively. （1） Sum of numbers tt （2） , will tt （1） With tt （2） The absolute value of the difference is used as the refresh interval of the jamming target; If it is less than 1, then calculate the average shooting interval ta of all trainees. （a1） and standard deviation ts （a1） ,exist Within the interval, a number is randomly selected as the refresh interval for the jamming target; Calculate the weighted effective scores of all trainees when shooting at fixed targets, and determine whether the weighted effective scores are qualified. If the results are unsatisfactory, then trainees with unsatisfactory results will be trained. If it passes the test, adjust the refresh interval of the dangerous target.

[0011] Furthermore, the process for adjusting the dangerous target refresh interval is as follows: Calculate the average number of hits (hn) for all trainees based on each trainee's hit rate and number of shots. （we） ; The hit rate of each trainee is calculated using a binomial distribution. （we） The probability of hitting the target each time is used to obtain ph. （1） ~ph （tn） Where tn represents the number of trainees; Calculate ph （1） ~ph （tn） mean ph （age） ; With a 5% margin of error, the hit rate is ph （age） The average shooting interval of the trainees is used as the shooting interval for dangerous targets; Calculate the weighted effective scores of all trainees when shooting at a fixed dangerous target, and determine whether the weighted effective scores are qualified. If the results are unsatisfactory, then trainees with unsatisfactory results will be trained. If successful, the moving target training submodule will be executed.

[0012] Furthermore, the workflow of the moving target training submodule is as follows: The moving target training submodule refreshes ordinary targets and interference targets around each trainee according to the interference target refresh interval set by the fixed target training submodule. Determine whether the valid scores of all trainees are satisfactory; If the test is not passed, training continues; If the training is successful, the shooting angle and effective score of each trainee are recorded, and the position and number of ordinary targets and interference targets are adjusted. Obtain the horizontal and vertical angles of the target when team member A hits it, and convert them into a direction vector to obtain ua. （1） ~ua （ns） Where ns represents the number of times team member A hits the target; Construct the standard direction vector: ; Fit probability density : ; in, Let T denote the normalization parameter, and T denote the transpose of the matrix. Indicates the smoothing parameter; Fitting The probability density corresponding to the target direction is obtained as follows: ; Obtain the calibration range of the probability density: [v （cal） v （ibr） ],Will Sort in ascending order and remove those that do not belong to [v] （cal） v （ibr） The values ​​within the specified interval are used to obtain the sequence fu; Extracting the first from sequence fu The number, used as a reference value v （val） Where α represents the score coefficient, and n （fu） This represents the number of values ​​in the sequence fu; The probability density is less than v （val） The standard direction serves as the shooting blind spot for Team A; Identify the shooting blind spots of other trainees and adjust the refresh interval of dangerous targets.

[0013] Furthermore, the workflow for adjusting the hazardous target refresh interval is as follows: Set the initial refresh interval for the dangerous targets, calculate the variance of the effective scores of all trainees, and obtain va. （1） ~va （tn） Where tn represents the number of trainees; Calculate the mean variance va （wei） ; If the variance of team member A's effective scores is va （甲） Then compare va（甲） With va （wei） The size determines the refresh interval of the dangerous targets around Team Member A; If va （甲） ≥va （wei） Then reduce the initial refresh interval. , which serves as the refresh interval for the dangerous target of Team Member A; If va （甲） <va （wei） Then increase the initial refresh interval. , which serves as the refresh interval for the dangerous target of Team Member A; Based on the refresh interval of the dangerous target for Team A, refresh one dangerous target in each of Team A's shooting blind spots in turn until the variance of Team A's effective score changes significantly or Team A is eliminated. When the variance of A's effective score changes significantly, stop refreshing dangerous targets around A, repeat the steps above to determine A's shooting blind zone and dangerous target refresh interval, and re-analyze and set A's shooting blind zone and dangerous target refresh interval. Adjust the refresh interval of dangerous targets for other trainees until all trainees' valid scores are qualified, then proceed to the mutual confrontation training submodule.

[0014] Furthermore, the workflow of the adversarial training submodule is as follows: Maximum number of advance targets n （adv） The maximum number of clusters of advance targets is set to n. （lim） ; Obtain the locations of all trainees who have not been eliminated, and determine whether all trainees who have not been eliminated are in the same area; If they are in the same area, count the total number of advance targets in that area; If the number of advance targets is < n （adv） If the number of advance targets reaches n, then advance targets will be sent to that area. （lim） ; If the number of advance targets = n （adv） If so, then do not process it; If the trainees who have not been eliminated are scattered in different areas, adjust the movement path of the advance target to the area where the trainees are gathered. Determine the path of the advance target to Cluster Area 1; Assuming there are ψ advance target deployment areas, calculate and determine the time cost from each deployment area to the first cluster area in turn, and select a path for the advance target. The deployment area of ​​the first advance target is designated as area A, and the total number of paths ro from area A to the first gathering area is obtained. On paths 1 to ro, obtain the movement speed vq of the advance target in each path. （1）~vq （ro） and the number of advance targets te （1） ~te （ro） Based on the path length from path 1 to path ro and the advance target's movement speed, calculate the arrival time for each path. Determine whether the movement speed of the advance target or the support target is the same on each path, and calculate the time cost of each path.

[0015] Furthermore, the process for selecting a path for the advance target is as follows: Scenario a: The advance target or the backup target moves at the same speed on each path; Calculate the comparison coefficients co for paths 1 to ro. （1） ~co （ro） ; Select co （1） ~co （ro） The path corresponding to the maximum value in the middle is taken as the target path of area A, and the arrival time corresponding to this path is taken as the time cost of area A. Scenario b: The movement speed of the advance target or the support target is different on each path; Get the current number of dangerous targets in cluster area 1 (tg) （now） ; Determine if path 1 is a candidate path, and calculate the minimum number of additional dispatches (tg). （1） : ; If te （1） ≥tg （1） If so, then path 1 will be selected as the alternative path; If te （1） <tg （1） If so, then do not process it; Similarly, determine whether path 2 to path ro is a candidate path; Among all the alternative paths, the path with the shortest arrival time is selected as the target path for area A, and the arrival time corresponding to this path is used as the time cost for area A. Determine the time cost of the 2nd to ψth expansion areas, take the expansion area with the minimum time cost as the starting area for expanding advance targets to cluster area 1, and take the path with the minimum time cost corresponding to this area as the path of advance target cluster area 1. The deployment path for advance targets is selected from other training personnel gathering areas.

[0016] Compared with the prior art, the beneficial effects of the present invention are: Achieving realistic shooting behavior simulation: This invention simulates a "curved bullet trajectory" based on the actual shooting state of the training environment and the trainees. It can mathematically model the real shooting direction data of the trainees under different postures, time periods, and shooting rhythms, thereby constructing a "bullet trajectory" that is close to the ballistic behavior of real firearms. This solves the problem that traditional laser simulation systems can only simulate the "geometric straight trajectory" of bullets under ideal conditions and cannot simulate the slight vibrations caused by the environment on the bullet trajectory.

[0017] Automatic identification of blind spots in shooting for trainees: This invention uses spherical density estimation and quantile thresholds to automatically identify blind spots in the shooting direction of trainees, eliminating reliance on manual analysis or subjective experience. Blind spots do not only refer to "directions that were not shot," but also directions with low statistical probability that are difficult for trainees to cover or where their reaction is slow. The system automatically adjusts the position of the enemy target based on the direction of the blind spot, forcing trainees to face their weaknesses and avoiding ineffective training caused by "practicing angles they are good at."

[0018] Automatic adjustment of training difficulty: By analyzing the shooting interval, hit rate, and number of times the target is hit by the trainees, the system can automatically calculate the target refresh rate and adaptively adjust the number of targets to keep the training difficulty fluctuating within a reasonable range and maintain the optimal training load. For example, the system will monitor the fluctuation of the trainees' effective performance in real time. When the fluctuation of the effective performance increases, the target refresh interval will be increased; when the fluctuation of the effective performance decreases, the target refresh interval will be shortened, so that the simulated training is neither too simple nor too difficult. Attached Figure Description

[0019] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the system modules of the present invention; Figure 2 This is a schematic diagram of the sub-module design of the present invention; Figure 3 This is a schematic diagram of the laser module of the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] Please see Figure 1 and Figure 3 The military training simulation system for AI-based adversarial technology based on reinforcement learning includes: Scene simulation module: used to simulate the flight speed and trajectory of flying objects (i.e., the "bullets" simulated by the laser module), and to construct differential equations 1 and 2; based on the activity status of the trainees and their actual positions, to set the response area of ​​the target or the response interval of the laser module. The targets used in this invention for counter-technology military training are divided into: ordinary targets, jamming targets, and dangerous targets; Ordinary target (i.e., a regular enemy target): Trainees can shoot at any position on an ordinary target and all scores will be counted as valid. Interference target (i.e., the enemy target holding hostages): Trainees only get a valid score if they shoot the enemy area within the interference target; if they shoot the hostage area within the interference target, their valid score will be deducted. Dangerous targets (i.e. enemy targets with weapons): Trainees' shots at any location on dangerous targets will be counted as valid scores; if a trainee is hit by a dangerous target, they will be immediately eliminated. The bullseye of the ordinary target, the jamming target, and the dangerous target are all located at the head position of the enemy in the corresponding area of ​​their respective targets; The workflow of the scene simulation module is as follows: Let the muzzle velocity of the projectile launcher (i.e., the laser module) be v. （o） The horizontal angle of the launch vehicle is a, and the elevation angle is b. (With the east direction as the positive x-axis, the north direction as the positive y-axis, and the vertical direction perpendicular to the ground as the positive z-axis) The initial velocities of the flying object along the x-axis, y-axis, and z-axis are respectively: vo （x） vo （y） vo （z） : ; Let the weight of the simulated projectile (i.e., the weight of the simulated projectile bullet) be m. （bul） Let the velocity of the flying object at time t be v. （t） :(vx （t） vy （t） vz （t） ), position is (x （t） y （t） , z （t） The wind speed is vw: (vw （x） vw （y） vw （z） ); The wind speed of the flying object is v （rel） :((vx （t） -vw （x） (vy) （t） -vw （y） (vz)（t） -vw （z） )); Construct the system of differential equations 1: ; System of differential equations 2: ; Where, ρ （air） C represents air density. D The drag coefficient is represented by g (determined by the "bullet shape" simulated by the laser module), g represents the acceleration due to gravity, and A represents the acceleration due to gravity. （rel） This represents the frontal area of ​​the "bullet" simulated by the laser module: ; With vo （x） vo （y） and vo （z） For velocity v （t） The initial value (i.e., vx) （t） vy （t） and vz （t） The initial values ​​are vo （x） vo （y） and vo （z） The flight speed and trajectory of the simulated flying object (i.e., the "bullet" simulated by the laser module) are determined based on differential equations 1 and 2. If the training area is indoors or in a windless area, the wind speed vw in differential equations 1 and 2 is set to zero (i.e., vw...). （x） vw （y） and vw （z） (Set the value to zero); Get the location of the trainees (x) （tr） y （tr） , z （tr） ) and bullseye position (x （ta） y （ta） , z （ta） ), obtain the wind speed (wx, wy, wz) of the training field; and set the target response time interval based on the activity status of the trainees and their actual positions. Scenario 1: Trainees remain stationary (i.e., trainees conduct shooting training in a fixed position); Based on the location of the trainees (x) （tr） y （tr） , z （tr） ) and bullseye position (x （ta） y （ta） , z （ta） ), calculate the actual distance l from the trainee to the target. （dis） Calculate the vertical distance l from the trainee to the target center. （ver） ; Determine if there is wind in the training area where the trainee is located; if not, do not set the target response area and the laser module response interval, and only detect whether the laser module operated by the trainee hits the bullseye. If so, the flight trajectory of the flying object is approximated by a parabolic trajectory, and the response interval of the target and the response interval of the laser module are set. Calculate the ideal flight time t of the flying object (i.e., the "bullet" simulated by the laser module). （ide） : ; Suppose that in the simulation training of scenario one, the horizontal angle at the time of launch of the aircraft is θ. （1） The elevation angle is φ （1） ; Construct equation 1-1: ; The horizontal angle θ can be obtained by solving equation 1-1. （1） and elevation angle φ （1） The mathematical expression: ; The target's response region is: ; The trainee's projectile launcher hit its target: The corresponding area is also considered as the bullseye for the trainees; Calculate the time t for the laser to reach the target. （las） : ; where c （光） Represents the speed of light; The response interval of the laser module is t. （del） : ; The delay t of the trainer's flight launcher (i.e., laser module) （del） The laser will fire after seconds; It should be noted that, since the present invention is based on using laser equipment to simulate the actual flight path of a "bullet", the situation described in the above "differential equation set 1 and differential equation set 2" is: the laser module simulates the flight path of the "bullet" under the influence of gravity, wind speed and air resistance during actual flight, that is, the non-parabolic "ballistic curve" in normal environment. Furthermore, since the present invention actually uses a laser module (using a "laser" to simulate a "bullet", and the laser propagates in a straight line), and the laser has no gravity or air resistance, in the design scenario of this application, only the influence of wind speed is considered when considering the response area of ​​the target or the response interval of the laser module. Scenario 2: Trainees can move (i.e., the trainees' positions are not fixed, and the trainees can choose their own shooting positions). Obtain the slowest reflex arc reaction time t of the trainee （ref） ; Determine whether the distance from the trainee to any target center is within the standard distance l （sta） Within; standard distance l （sta） : ; If a trainee is within the standard distance l from the bullseye of the target closest to them. （sta） Within this range, no target response area or laser module response interval is set; only the laser module operated by the trainee is tested to see if it hits the bullseye. If the distance between the trainee and the bullseye of the nearest target is not within the standard distance... （sta） If the distance is within 1 lb, then calculate the distance lb from the trainee to the bullseye. （tar） ; The response interval of the laser module is ti. （tar） : ; In scenario two, no target response area is set.

[0022] Training simulation module: used to acquire the position of trainees in real time, and combine differential equations 1 and 2 with the response area of ​​the target or the response interval of the laser module to count and record the shooting interval and effective score of each trainee; based on the shooting interval and effective score of the entire training, the fixed target training submodule, moving target training submodule and mutual confrontation training submodule are executed in sequence. Please see Figure 2 The training simulation module includes: a fixed target training submodule, a moving target training submodule, and a mutual adversarial training submodule. Fixed Target Training Submodule: Used to acquire and analyze the shooting interval of trainees shooting at ordinary targets, and adjust the refresh interval of interference targets; acquire and analyze the number of shots, shooting interval and hit rate of trainees shooting at interference targets, adjust the shooting interval of dangerous targets, and determine whether the effective scores of all trainees are qualified; if qualified, the moving target training submodule is executed; if unqualified, training continues. (The fixed target training submodule uses ordinary targets, interference targets, and dangerous targets.) The workflow of the fixed target training submodule is as follows: Based on differential equations 1 and 2 and the response region of the target or the response interval of the laser module, the shooting interval and effective score of each trainee shooting at a normal target are statistically analyzed and recorded. The total number of trainees was counted, along with the number of trainees with valid passing scores and the number of trainees with valid failing scores. The percentage of trainees who achieved a passing score (pr) during the fixed-target shooting exercise was then calculated. （a1）and the percentage of students who failed the test （a2） ; Calculate the average effective score (gr) of all trainees who achieved a passing score during fixed-target shooting practice. （a1） The average valid score (gr) of all trainees who achieved a passing grade. （a2） Calculate the weighted effective score wg of all trainees during fixed-target shooting practice. （a） : ; The weighted effective score (wg) of all trainees during fixed-target shooting is determined. （a2） Is it qualified? If the score is not satisfactory, the trainees who fail to achieve a satisfactory score during the fixed-target shooting exercise will be trained until they achieve a satisfactory score. If the training is successful, the shooting interval and hit rate of each trainee will be recorded, and the refresh interval of the interference target will be adjusted. Standardize the firing interval and hit rate of all trainees (e.g., Z-Score standardization), and then calculate the correlation coefficient between the (standardized) firing interval and the (standardized) hit rate. Determine if the absolute value of the correlation coefficient is greater than 0.8; If it is greater than 1, then a multinomial regression equation is constructed with the shooting interval of all trainees as the independent variable and the hit rate of all trainees as the dependent variable. (Using the hit rate of all trainees shooting at a fixed target as the domain of the polynomial regression equation) Calculate the numbers tt that minimize and maximize the polynomial regression equation. （1） Sum of numbers tt （2） , as the refresh interval of the interference target; If there are multiple numbers that minimize or maximize the polynomial regression equation, then remove the maximum and minimum values ​​of these numbers, calculate the average of the remaining numbers, and then... （1） With tt （2） The absolute value of the difference is used as the refresh interval of the jamming target; If it is less than ta, then calculate the average shooting interval ta of all trainees. （a1） and standard deviation ts （a1） ,exist Within the interval, a number is randomly selected as the refresh interval for the jamming target; Repeat the above weighted effective score wg （a） The calculation process involves calculating the weighted effective scores of all trainees when shooting at fixed targets, and determining whether the weighted effective scores of all trainees when shooting at fixed targets are qualified. If the test is not passed, the trainees who fail to achieve a passing score during fixed-target shooting will be trained until they achieve a passing score. If the training is successful, the number of shots, shooting intervals, and hit rates of each trainee will be recorded, and the refresh interval of the dangerous targets will be adjusted accordingly. Calculate the average number of hits (hn) for all trainees based on each trainee's hit rate and number of shots. （we） ; The hit rate of each trainee is calculated using a binomial distribution. （we） The expected hit probability of each number is given by ph. （1） ~ph （tn） Where tn represents the number of trainees; The calculation process for the (expected) hit probability is as follows: Let h be the number of shots fired by a trainee. （x） The hit rate is p （x） The trainee hit hn （we） The probability of this is P(hn) （we） ): ; Calculate ph （1） ~ph （tn） mean ph （age） ; With a 5% margin of error, the hit rate is ph （age） The average shooting interval of the trainees is used as the shooting interval for dangerous targets; Repeat the above weighted effective score wg （a） The calculation process involves calculating the weighted effective scores of all trainees when shooting at a fixed dangerous target, and determining whether the weighted effective scores of all trainees when shooting at a fixed dangerous target are qualified. If the test is not passed, the trainees who fail to achieve a passing score during fixed-target shooting will be trained until they achieve a passing score. If successful, the moving target training submodule will be executed. Moving Target Training Submodule: This module randomly refreshes ordinary and distracting targets around the trainees and determines whether the effective scores of all trainees are qualified. If they are not qualified, training continues. If they are qualified, dangerous targets are refreshed around the trainees based on their shooting angles and effective scores. The effective scores of all trainees are then tallied and determined again. If they are qualified, the mutual confrontation training submodule is executed. If they are not qualified, training continues. (The moving target training submodule uses ordinary targets, jamming targets, and dangerous targets.) The workflow of the moving target training submodule is as follows: The moving target training submodule refreshes ordinary targets and interference targets around each trainee according to the interference target refresh interval set by the fixed target training submodule. Determine whether the valid scores of all trainees are satisfactory; If the test is not passed, training continues; If qualified, the blind spots and fluctuations in the effective scores of each trainee will be analyzed based on their shooting angles and effective scores, and dangerous targets will be refreshed around the trainee. Analyze the shooting angles of trainees to determine blind spots; Record the shooting angle and effective score of each trainee when hitting the target, analyze the shooting angle of each trainee, and adjust the position and refresh number of ordinary targets and interference targets; Team member A was the trainee participating in this moving target training. Obtain the horizontal angle θa of each time team member A hits the target. （1） θa （2） ~θa （ns） and elevation angle φa （1） ,φa （2） ~φa （ns） Where ns represents the number of times team member A hits the target; Where, θa （1） ~θa （ns） The range of values ​​for is: [-180] o 180 o ];φa （1） ~φa （ns） The range of values ​​for is: [-90] o 90 o ]; Convert the horizontal and vertical angles of soldier A's firing position into direction vectors to obtain ua. （1） ua （2） ~ua （ns） ; Among them, ua （1） The mathematical expression is: ; ua （2） The mathematical expression is: ; And so on, ua （ns） The mathematical expression is: ; The aforementioned direction vector ua （1） ~ua （ns） They are all in (3×1) matrix format; With ua （1）Based on this, construct a standard direction vector: ; in, The mathematical expression is: ; The mathematical expression is: ; Similarly, The mathematical expression is: ; Based on the "spherical kernel density estimation (i.e., the von Mises–Fisher kernel function)," the shooting of Team A was fitted. Corresponding to the target direction (i.e., the horizontal angle is -180°) o Angle of elevation is -90 o probability density : ; in, Indicates the normalization parameter: ; This represents matrix multiplication, and T represents the transpose of the matrix; This represents the smoothing parameter (used to control the smoothness of the von Mises–Fisher kernel function). The suggested values ​​are shown in the table below: Reference table for possible values; repeat The calculation process, fitting the shot of player A. The probability density corresponding to the target direction is obtained. ; Obtain the calibration range of the probability density: [v （cal） v （ibr） ],Will Sort in ascending order and remove those that do not belong to [v] （cal） v （ibr） The values ​​within the specified interval are used to obtain the sequence fu; Extracting the first from sequence fu The number of blind spots for determining the shooting blind zone of Team A is used as a reference value v. （val） Where α represents the score coefficient (α ranges from 0 to 1), n （fu） This represents the number of values ​​in the sequence fu; like If it is not an integer, then extract. The integer part i in（α） For the i-th sequence fu （α） The and the (i) （α） +1) count, perform linear interpolation to obtain the reference value v （val） ; The probability density is less than v （val） The standard direction serves as the shooting blind spot for Team A; Example explanation: Assume it belongs to [v] （cal） v （ibr） There are 10 values ​​in the range: The sequence fu is: 0.01, 0.02, 0.03, 0.04, 0.05, 0.07, 0.08, 0.09, 0.10, 0.12; Scenario 1: If the value of α is 0.1, that is The value is an integer 2: ; v （val） The value is: the second value in the sequence fu, which is 0.02; There is only one value with a probability density less than 0.02: 0.01. The shooting blind zone is (90°, 20°), meaning that trainees lack shooting opportunities at positions with a horizontal angle of 90° and an elevation angle of 20°. Scenario 2: If the value of α is 0.35, that is Not an integer: ; The integer part is 3, that is: ; The third and fourth values ​​in sequence fu are 0.03 and 0.04, respectively. Linear interpolation of 0.03 and 0.04 yields v. （val） The value is: 0.035; v （val） The calculation formula is: ; The probability density values ​​less than 0.035 are 0.01, 0.02, and 0.03. The blind spots for firing are: (90°, 20°), (30°, 10°), and (60°, 10°). That is, the trainees lacked shooting skills at positions with a horizontal angle of 90° and an elevation angle of 20°, a horizontal angle of 30° and an elevation angle of 10°, and a horizontal angle of 60° and an elevation angle of 10°. Repeat the steps to determine the shooting blind spots of Team Member A, and determine the shooting blind spots of other trainees "during the moving target training process"; Analyze the effective performance of trainees and adjust the refresh interval of dangerous targets; Set the initial refresh interval for dangerous targets (i.e., the refresh interval for ordinary targets and interference targets set in the "Moving Target Training Submodule"). Calculate the variance of the effective scores of all trainees to obtain va. （1） ~va （tn） Where tn represents the number of trainees; Extracted maximum value va （min） and minimum value va （min） Calculate the (weighted) average variance va （wei） : ; Among them, va （s） Let represent the variance of the effective performance of the s-th trainee, where s ranges from 1 to tn. If the variance of team member A's effective scores is va （甲） Then compare va （甲） With va （wei） The size determines the refresh interval of the dangerous targets around Team Member A; If va （甲） ≥va （wei） This indicates that the fluctuation in Team A's effective score is relatively large, so the initial refresh interval should be reduced. , which serves as the refresh interval for the dangerous target of Team Member A; If va （甲） <va （wei） This indicates that the fluctuation of Team A's effective score is relatively small, and the initial refresh interval should be increased. , which serves as the refresh interval for the dangerous target of Team Member A; Based on the refresh interval of the dangerous target for Team A, refresh one dangerous target in each of Team A's shooting blind spots in turn until the variance of Team A's effective score changes significantly or Team A is eliminated. When the variance of A's effective score changes significantly, stop refreshing dangerous targets around A, repeat the steps above to determine A's shooting blind zone and dangerous target refresh interval, and re-analyze and set A's shooting blind zone and dangerous target refresh interval. It should be noted that "significant changes in the variance of Team A's effective scores" means that "the current variance of Team A's effective scores" has increased or decreased by at least 30% compared to "the variance of Team A's effective scores in the previous period". Here, "30%" is just a reference coefficient set by this invention to judge the change in variance. Users or relevant technical personnel can adjust the size of the reference coefficient according to actual needs. Repeat the same steps to adjust the dangerous target refresh interval for Team A, and adjust the dangerous target refresh interval for other trainees "participating in moving target training" until all trainees' valid scores are qualified, then enter the mutual confrontation training sub-module. The mutual confrontation training submodule is used to divide dangerous targets into advance targets and support targets, and set the maximum number of advance targets to be deployed and the maximum number of targets to be gathered; adjust the movement paths of advance targets and support targets according to the location of the trainees; count the effective scores of all trainees, and retrain the trainees with unsatisfactory effective scores until they pass the effective score. (The adversarial training submodule only uses dangerous targets.) The workflow of the adversarial training submodule is as follows: Let n be the maximum number of advance targets to be deployed. （adv） The maximum number of clusters of advance targets is set to n. （lim） ; "Maximum number of advance targets" refers to the total number of advance targets allowed in the entire training area; "Maximum number of advance targets" means: the maximum number of advance targets allowed in a single area of ​​the training region; (For example: the training area has three areas, A, B, and C; "maximum number of advance targets" refers to the total number of advance targets in the three areas A, B, and C; "maximum number of advance targets clustered" refers to the total number of advance targets in one of the areas A, B, or C.) Obtain the locations of all trainees who have not been eliminated, and determine whether all trainees who have not been eliminated are in the same area; If they are in the same area, count the total number of advance targets in that area; If the number of advance targets is < n （adv） If n targets are found, additional advance targets will be added to the "area where all trainees are located" until the number of advance targets in the "area where all trainees are located" reaches n. （lim） One, or send out all backup targets; If the number of advance targets = n （adv） If there are only a few, then no action will be taken; If the trainees who have not been eliminated are scattered in different areas, then the movement path of the advance target is adjusted according to the number of trainees who have not been eliminated in each area, and in combination with the movement speed of the advance target and the backup target. Count the number of trainees who were not eliminated in each region; Based on the descending order of the number of trainees who have not been eliminated in each area, the areas where the trainees are located are arranged as follows: Area 1 (trainees), Area 2 (trainees), and Area ω (trainees). ω represents the total number of regions containing all trainees who were not eliminated. Determine the path for the advance target to the No. 1 (trainee) assembly area; Assuming there are ψ advance target reinforcement areas, calculate and determine the time cost from each reinforcement area to the No. 1 (trainee) gathering area in turn, and select a path for the advance target. The deployment area of ​​the first advance target is designated as area A. The total number of paths ro from area A to the gathering area of ​​trainee number 1 is obtained. On path 1, path 2, and up to path ro, obtain the movement speed vq of the advance target in each path. （1） vq （2） ~vq （ro） and the number of advance targets te （1） te （2） ~te （ro） ; Based on the path lengths of paths 1 and 2 up to path ro and the movement speed of the advance target, calculate the arrival time ar from area A to the training personnel's assembly area 1 for each path. （1） ar （2） ~ar （ro） ; (On path 1 to path ro) Determine whether the movement speed of the advance target or the support target is the same on each path, and calculate the time cost of each path; Scenario a: (On path 1 to path ro) The advance target or the backup target moves at the same speed on each path; For te （1） te （2） ~te （ro） and ar （1） ar （2） ~ar （ro） Z-score normalization yields the dimensionless number of ze. （1） ,ze （2） ~ze （ro） and dimensionless time zr （1） ,zr （2） ~zr （ro）； Calculate the comparison coefficient co for the first path. （1） : ; The comparison coefficient co for the second path （2） : ; And so on, the comparison coefficient co for the ro-th path （ro） : ; Select co（1） ~co （ro） The path corresponding to the maximum value in the middle is taken as the target path of area A, and the arrival time corresponding to this path is taken as the time cost of area A. Scenario b: (On path 1 to path ro) The movement speed of the advance target or the support target is different on each path; The target hit frequency fb in the No. 1 (trainee) gathering area is calculated (i.e., the average number of dangerous targets hit per minute by non-eliminated trainees in the No. 1 (trainee) gathering area, in units of targets / minute). When dispatching advance targets from support target path 1 to path ro, determine whether they pass through area A, and calculate the time cost of area A; Branch b1: When the backup target is sent to the No. 1 (trainee) gathering area via path 1 to path ro, it always passes through area A (the movement speed of the backup target on each path is not considered). Get the current number of dangerous targets (tg) within the gathering area of ​​trainee #1. （now） ; Determine if path 1 is a candidate path, and calculate the minimum number of advance targets tg for path 1. （1） : ; If the number of advance targets on path 1 is te （1） ≥tg （1） If so, then path 1 will be selected as the alternative path; If the number of advance targets on path 1 is te （1） <tg （1） If so, then do not process it; Determine if path 2 is a candidate path, and calculate the minimum number of advance targets tg for path 2. （2） : ; If the number of advance targets on path 1 is te （2） ≥tg （2） If so, then path 2 will be selected as the alternative path; If the number of advance targets on path 1 is te （2） <tg （2） If so, then do not process it; Similarly, determine whether path ro is a candidate path, and calculate the minimum number of advance targets tg for path ro. （ro） : ; If the number of advance targets on path 1 is te （ro） ≥tg （ro） Then the path ro As an alternative route; If the number of advance targets on path 1 is te （ro） <tg （ro） If so, then do not process it; Among all the alternative paths, the path with the shortest arrival time is selected as the target path for area A, and the arrival time corresponding to this path is used as the time cost for area A. Branch b1 simulates a combat situation where "the enemy has supplies" (i.e., there is a supply line from "support target to area A" in area A). Branch b2: When the backup target is dispatched to the No. 1 (trainee) gathering area via path 1 to path ro, it does not pass through area A; Get the length L of path 1 （1） And the movement speed vq of the advance target on path 1. （1） and the movement speed vh of the backup target （1） ; Obtain the number of advance targets nq in path 1 （1） and the number of backup targets nh （1） ; Get the current number of dangerous targets (tg) within the gathering area of ​​trainee #1. （now） And the target hit frequency fb in the area where trainees gather (1), then determine whether the target update is round trip, and construct a mathematical model for path 1 (the number of targets changes over time): Branch b21: The advance target and the backup target are updated in a round-trip manner (that is, after the "advance target and backup target" reach the No. 1 (trainee) gathering area, the "No. 1 (trainee) gathering area" returns an equal number of hit targets, and if there are not enough, all targets are returned). A single advance target travels along path 1 to the training personnel's assembly area 1 and returns within a cycle T of a hit target. （a） : ; where v （rev） Indicates the return speed of a hit target; A single backup target travels along path 1 to the training personnel's assembly area and returns within a cycle T of a hit target. （b） : ; At time t, the number of advance targets reaching the No. 1 (trainee) gathering area is Na. （t） : ; At time t, the number of support targets arriving at the training personnel's (Nb) assembly area is Nb. （t） : ; The mathematical model for path 1 (the number of targets changing over time) is Ir（t） : ; Branch b21 simulates a combat scenario where the "enemy is being treated," and is suitable for situations with a small number of targets or a large number of trainees. Branch b22: Advance targets and backup targets are updated in a one-way manner (i.e., after the "advance targets and backup targets" reach the No. 1 (trainees) gathering area, targets that have not been hit return), which is suitable for situations with a large number of targets or a small number of trainees; Redefine the number of advance targets (Trainees) arriving at the gathering area of ​​Personnel 1 at time t, Ua （t） : ; Redefine the number of backup targets (Ub) arriving at the training personnel's (1) gathering area at time t. （t） : ; The mathematical model for path 1 (the number of targets changing over time) is Ia. （t） : ; Branch b22 simulates a combat scenario where the enemy has not been healed; Branch b2 simulates a combat situation where "the enemy has no supplies" (i.e., there is no supply line from "support target to area A" in area A). Repeat the same steps as constructing the mathematical model of path 1 (the number of targets changes over time) to construct the mathematical model of path 2 to path ro (the number of targets changes over time); based on the mathematical model of path 1 to path ro (the number of targets changes over time), find the path with the largest increase in the number of targets per unit time, take it as the target path of area A, and take the arrival time corresponding to this path as the time cost of area A. Branch b3: When the backup target is dispatched to the No. 1 (trainee) gathering area via path 1 to path ro, some targets pass through area A and some do not pass through area A. Based on whether the backup targets pass through Area A (i.e., whether the advance targets are sent to the No. 1 (trainee) assembly area via path 1 to path ro, the path 1 to path ro is divided into a set Z that passes through Area A. （1） and sets that do not pass through area A （2） ; Based on the selection strategy of branch b1, in set Z （1） Select the target path Za in area A; Based on the selection strategy of branch b2, in set Z （2） Select the target path Zb in area A; Among the target paths Za and Zb, the path with the minimum time cost is selected as the target path for area A, and the arrival time of this path is taken as the time cost of area A. If the time costs of target path Za and target path Zb are the same, then choose randomly; Repeat the steps to determine the time cost of area A, determine the time cost of the 2nd to ψth expansion areas (i.e. advance target expansion areas), take the expansion area with the smallest time cost as the starting area for expanding advance targets to the No. 1 (trainee) gathering area, and take the path with the smallest time cost corresponding to this area as the path of advance target No. 1 (trainee) gathering area. Repeat the above steps to select advance target deployment paths for the gathering areas of trainees from 2 to ω in turn; If the deployment paths of two or more (trainees) clusters overlap, the deployment path will be assigned to the "(trainees) cluster with a large number of trainees, high effective scores, and low variance of effective scores". For example, the deployment path for gathering areas 1, 2, and 3 (trainees) is path C; Situation B （1） Compare the number of trainees in clusters 1, 2, and 3 (trainees), and assign "Path C" to the cluster with the largest number of trainees. Situation B （2） If situation B （1） If the attribution of "Path C" cannot be determined, then compare the total effective scores of trainees in clusters 1, 2, and 3 (trainees) and assign "Path C" to the cluster with the largest total effective score. Situation B （3） If situation B （2） If the attribution of "Path C" still cannot be determined, then compare the variance of the effective performance of trainees in clusters 1, 2 and 3 (trainees) and assign "Path C" to the cluster with the smallest variance of effective performance. If scenario B （3） If the attribution of "Path C" still cannot be determined, then "Path C" will be randomly assigned to one of the "1, 2 and 3 (trainees) gathering areas".

[0023] Performance Feedback Module: Used to statistically analyze and record the effective performance of all trainees, summarize the information of trainees who participated in intensive combat training, and provide feedback.

[0024] The above formulas are all dimensionless calculations. The formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation. For example, there are weighting coefficients and proportional coefficients. The values ​​set are to quantify each parameter to obtain a specific value, which is convenient for subsequent comparison. The values ​​of the weighting coefficients and proportional coefficients are only required to not affect the proportional relationship between the parameters and the quantified values.

[0025] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A military training simulation system for artificial intelligence adversarial technology based on reinforcement learning, characterized in that, include: Scene simulation module: used to simulate the flight speed and trajectory of flying objects and construct a system of differential equations; Based on the trainees' activity status and actual location, set the target's response area or the laser module's response interval. Training simulation module: used to acquire the position of trainees in real time, and combine the differential equation system and the response area of ​​the target or the response interval of the laser module to count and record the shooting interval and effective score of each trainee; based on the shooting interval and effective score of all trainees, adjust the refresh interval, refresh position and movement path of the target. Performance Feedback Module: Used to statistically analyze and record the effective performance of all trainees, summarize the information of trainees who participated in intensive combat training, and provide feedback.

2. The military training simulation system for artificial intelligence adversarial technology based on reinforcement learning according to claim 1, characterized in that, The training simulation module includes the following sub-modules: Fixed Target Training Submodule: Used to acquire and analyze the shooting interval of trainees shooting at ordinary targets, and adjust the refresh interval of interference targets; acquire and analyze the number of shots, shooting interval and hit rate of trainees shooting at interference targets, adjust the shooting interval of dangerous targets, and determine whether the effective scores of all trainees are qualified; if qualified, the moving target training submodule is executed; if unqualified, training continues. Moving Target Training Submodule: This module randomly refreshes ordinary and distracting targets around the trainees and determines whether the effective scores of all trainees are qualified. If they are not qualified, training continues. If they are qualified, dangerous targets are refreshed around the trainees based on their shooting angles and effective scores. The effective scores of all trainees are then tallied and determined again. If they are qualified, the mutual confrontation training submodule is executed. If they are not qualified, training continues. The mutual confrontation training submodule is used to divide dangerous targets into advance targets and support targets, and set the maximum number of advance targets to be deployed and the maximum number of targets to be gathered; adjust the movement paths of advance targets and support targets according to the location of the trainees; count the effective scores of all trainees, and conduct confrontation training again for trainees with unsatisfactory effective scores until the effective scores are satisfactory.

3. The military training simulation system for artificial intelligence adversarial technology based on reinforcement learning according to claim 2, characterized in that, The workflow of the scene simulation module is as follows: Based on the speed v of the launch device （o） Based on the horizontal angle, elevation angle, and the weight of the simulated flying object, construct differential equations 1 and 2; Obtain the location of the trainees and the target, and the wind speed in the training field; based on the trainees' activity status and actual location, set the target response time interval; Scenario 1: The trainees remain stationary; Calculate the actual distance l from the trainee to the target based on the trainee's position and the target's position. （dis） Calculate the vertical distance l from the trainee to the target center. （ver） ; Determine if there is wind in the training area where the trainees are located; If not, then the target response area and the laser module response interval are not set; If so, set the target response interval and the laser module response interval; Calculate the ideal flight time t of the flying object （ide） ; Get the horizontal angle θ （1） and elevation angle φ （1） Set the target's response area; Calculate the time t for the laser to reach the target. （las） ; The response interval of the laser module is t. （del） : ; Scenario 2: The trainees can move.

4. The military training simulation system for artificial intelligence adversarial technology based on reinforcement learning according to claim 3, characterized in that, The target response interval for scenario two is set as follows: Obtain the slowest reflex arc reaction time t of the trainee （ref） ; Determine whether the distance from the trainee to any target center is within the standard distance l （sta） Within; standard distance l （sta） : ; If a trainee's distance to the nearest target's bullseye is within the standard distance l （sta） Within this range, no target response area or laser module response interval is set; only the laser module operated by the trainee is tested to see if it hits the bullseye. If the distance between the trainee and the nearest target bullseye is not within the standard distance... （sta） If the response interval ti of the laser module is within a certain range, then calculate the response interval ti of the laser module. （tar） ; Scenario 2: No target response area is set; Based on differential equations 1 and 2 and the response region of the target or the response interval of the laser module, the shooting interval and effective score of each trainee shooting at a normal target are statistically recorded, and the fixed target training sub-module is executed.

5. The military training simulation system for artificial intelligence adversarial technology based on reinforcement learning according to claim 4, characterized in that, The workflow of the fixed target training submodule is as follows: Calculate the percentage of trainees who passed and failed, and then calculate the weighted average score (wg). （a） ; The weighted effective score (wg) of all trainees during fixed-target shooting is determined. （a2） Is it qualified? If the results are unsatisfactory, then trainees with unsatisfactory results will be trained. If it passes the test, adjust the refresh interval of the jamming target. The firing interval and hit rate of all trainees were standardized, and then the correlation coefficient between firing interval and hit rate was calculated. Determine if the absolute value of the correlation coefficient is greater than 0.8; If it is greater than 1, then a multinomial regression equation is constructed with the shooting interval of all trainees as the independent variable and the hit rate of all trainees as the dependent variable. Calculate the numbers tt that minimize and maximize the polynomial regression equation, respectively. （1） Sum of numbers tt （2） , will tt （1） With tt （2） The absolute value of the difference is used as the refresh interval of the jamming target; If it is less than ta, then calculate the average shooting interval ta of all trainees. （a1） and standard deviation ts （a1） ,exist Within the interval, a number is randomly selected as the refresh interval for the jamming target; Calculate the weighted effective scores of all trainees when shooting at fixed targets, and determine whether the weighted effective scores are qualified. If the results are unsatisfactory, then trainees with unsatisfactory results will be trained. If it passes the test, adjust the refresh interval of the dangerous target.

6. The military training simulation system for artificial intelligence adversarial technology based on reinforcement learning according to claim 5, characterized in that, The procedure for adjusting the refresh interval of dangerous targets is as follows: Calculate the average number of hits (hn) for all trainees based on each trainee's hit rate and number of shots. （we） ; The hit rate of each trainee is calculated using a binomial distribution. （we） The probability of hitting the target each time is used to obtain ph. （1） ~ph （tn） Where tn represents the number of trainees; Calculate ph （1） ~ph （tn） mean ph （age） ; With a 5% margin of error, the hit rate is ph （age） The average shooting interval of the trainees is used as the shooting interval for dangerous targets; Calculate the weighted effective scores of all trainees when shooting at a fixed dangerous target, and determine whether the weighted effective scores are qualified. If the results are unsatisfactory, then trainees with unsatisfactory results will be trained. If successful, the moving target training submodule will be executed.

7. The military training simulation system for artificial intelligence adversarial technology based on reinforcement learning according to claim 6, characterized in that, The workflow of the moving target training submodule is as follows: The moving target training submodule refreshes ordinary targets and interference targets around each trainee according to the interference target refresh interval set by the fixed target training submodule. Determine whether the valid scores of all trainees are satisfactory; If the test is not passed, training continues; If the training is successful, the shooting angle and effective score of each trainee are recorded, and the position and number of ordinary targets and interference targets are adjusted. Obtain the horizontal and vertical angles of the target when team member A hits it, and convert them into a direction vector to obtain ua. （1） ~ua （ns） Where ns represents the number of times team member A hits the target; Construct the standard direction vector: Fitted probability density: ； Obtain the calibration range of the probability density: [v （cal） v （ibr） ],Will Sort in ascending order and remove those that do not belong to [v] （cal） v （ibr） The values ​​within the specified interval are used to obtain the sequence fu; Extracting the first from sequence fu The number, used as a reference value v （val） Where α represents the score coefficient, and n （fu） This represents the number of values ​​in the sequence fu; The probability density is less than v （val） The standard direction serves as the shooting blind spot for Team A; Identify the shooting blind spots of other trainees and adjust the refresh interval of dangerous targets.

8. The military training simulation system for artificial intelligence adversarial technology based on reinforcement learning according to claim 7, characterized in that, The workflow for adjusting the refresh interval of hazardous targets is as follows: Set the initial refresh interval for the dangerous targets, calculate the variance of the effective scores of all trainees, and obtain va. （1） ~va （tn） Where tn represents the number of trainees; Calculate the mean variance va （wei） ; If the variance of team member A's effective scores is va （甲） Then compare va （甲） With va （wei） The size determines the refresh interval of the dangerous targets around Team Member A; If va （甲） ≥va （wei） This shortens the initial refresh interval; If va （甲） <va （wei） If so, the initial refresh interval will be extended; Based on the refresh interval of the dangerous target for Team A, refresh one dangerous target in each of Team A's shooting blind spots in turn until the variance of Team A's effective score changes significantly or Team A is eliminated. When the variance of A's effective score changes, stop refreshing dangerous targets around A, repeat the steps above to determine A's shooting blind zone and dangerous target refresh interval, and re-analyze and set A's shooting blind zone and dangerous target refresh interval. Adjust the refresh interval of dangerous targets for other trainees until all trainees' valid scores are qualified, then proceed to the mutual confrontation training submodule.

9. The military training simulation system for artificial intelligence adversarial technology based on reinforcement learning according to claim 2, characterized in that, The workflow of the adversarial training submodule is as follows: Maximum number of advance targets n （adv） The maximum number of clusters of advance targets is set to n. （lim） ; Obtain the locations of all trainees who have not been eliminated, and determine whether all trainees who have not been eliminated are in the same area; If they are in the same area, count the total number of advance targets in that area; If the number of advance targets is < n （adv） If the number of advance targets reaches n, then advance targets will be sent to that area. （lim） ; If the number of advance targets = n （adv） If so, then no action will be taken; If the trainees who have not been eliminated are scattered in different areas, adjust the movement path of the advance target to the area where the trainees are gathered. Determine the path of the advance target to Cluster Area 1; Assuming there are ψ advance target deployment areas, calculate and determine the time cost from each deployment area to the first cluster area in turn, and select a path for the advance target. The deployment area of ​​the first advance target is designated as area A, and the total number of paths ro from area A to the first gathering area is obtained. On paths 1 to ro, obtain the movement speed vq of the advance target in each path. （1） ~vq （ro） and the number of advance targets te （1） ~te （ro） Based on the path length from path 1 to path ro and the advance target's movement speed, calculate the arrival time for each path. Determine whether the movement speed of the advance target or the support target is the same on each path, and calculate the time cost of each path.

10. The military training simulation system for artificial intelligence adversarial technology based on reinforcement learning according to claim 9, characterized in that, The process for selecting a path for the advance target is as follows: Scenario a: The advance target or the backup target moves at the same speed on each path; Calculate the comparison coefficients co for paths 1 to ro. （1） ~co （ro） ; Select co （1） ~co （ro） The path corresponding to the maximum value in the middle is taken as the target path of area A, and the arrival time corresponding to this path is taken as the time cost of area A. Scenario b: The movement speed of the advance target or the support target is different on each path; Get the current number of dangerous targets in cluster area 1 (tg) （now） ; Determine if path 1 is a candidate path, and calculate the minimum number of additional dispatches (tg). （1） : ； If te （1） ≥tg （1） If so, then path 1 will be selected as the alternative path; If te （1） <tg （1） If so, then no action will be taken; Similarly, determine whether path 2 to path ro is a candidate path; Among all the alternative paths, the path with the shortest arrival time is selected as the target path for area A, and the arrival time corresponding to this path is used as the time cost for area A. Determine the time cost of the 2nd to ψth expansion areas, take the expansion area with the minimum time cost as the starting area for expanding advance targets to cluster area 1, and take the path with the minimum time cost corresponding to this area as the path of advance target cluster area 1. The deployment path for advance targets is selected from other training personnel gathering areas.

Images