A radar anti-jamming decision model evaluation method based on AHC

Abstract

This invention belongs to the field of radar anti-jamming strategy generation, specifically providing an evaluation method for radar anti-jamming decision models based on AHC (Adaptive Harmony Conversion), addressing the problems of static, single-dimensional, and insufficient robustness considerations in existing evaluation systems. This invention constructs a three-level structure—target layer, criterion layer, and indicator layer—covering the three core dimensions of effectiveness, efficiency, and robustness. It adds indicators such as interference type adaptability and training stability, overcoming the single-dimensionality deficiency of existing evaluation systems and achieving a comprehensive evaluation of anti-jamming strategies. Furthermore, by combining the AHP (Adaptive Hierarchy Process) method with a dynamic weight adjustment mechanism, the weights of each criterion layer are adaptively adjusted according to the interference intensity, enabling the evaluation system to adapt to different dynamic interference scenarios and improving the relevance and accuracy of the evaluation results. Based on this, this invention can achieve accurate and comprehensive evaluation of radar network anti-jamming strategies.

Description

Technical Field

[0001] This invention belongs to the field of radar anti-jamming strategy generation, specifically providing a radar anti-jamming decision model evaluation method based on AHC (Analytic Hierarchy Process). Background Technology

[0002] With the rapid development of electronic jamming technology, radars face increasingly complex jamming environments, making radar networks, with their multi-node collaborative advantages, an important direction for anti-jamming technology development. Currently, the evaluation of radar network anti-jamming strategies faces several problems: First, existing evaluation indicators mostly focus on the performance of individual radars, lacking quantification of the collaborative effects of multiple nodes; second, robustness indicators are insufficiently covered, failing to fully consider the impact of dynamic changes in jamming parameters; third, some indicator calculations rely on ideal data, ignoring the impact of jamming parameter estimation errors in actual combat; fourth, evaluation systems are mostly based on static weight allocation, unable to adapt to dynamic scenarios with different jamming intensities. These problems make it difficult for evaluation results to comprehensively and accurately reflect the actual effectiveness of anti-jamming strategies, and cannot effectively support strategy optimization and iteration.

[0003] In recent years, researchers' exploration of anti-jamming assessment indicators has evolved from single performance indicators to multi-dimensional comprehensive systems. The core trajectory can be divided into three stages: early research (focusing on single indicators of core effectiveness), mid-term research (moving towards multi-dimensional comprehensive assessment), and recent research (robustness assessment in the context of intelligent networking). The focus of indicator design at each stage is closely related to the technological background. Before 2000, radar and jamming technologies were relatively simple, and assessment indicators mainly revolved around the radar's core detection function, focusing on single, intuitive effectiveness indicators to quickly determine whether jamming was effectively suppressed. With the diversification of jamming technologies (the emergence of deceptive jamming) and the increasing complexity of radar systems, single indicators could no longer comprehensively measure anti-jamming capabilities. Between 2000 and 2015, a two-dimensional assessment framework combining effectiveness and efficiency began to be constructed. For multi-node collaborative anti-jamming, indicators such as node fault tolerance and collaborative gain coefficient were proposed to evaluate the network's redundancy capabilities and collaborative operational advantages when some nodes fail.

[0004] In summary, regardless of the historical development of the evaluation system, based on practical needs, the indicators can be divided into three main categories: effectiveness indicators (anti-jamming effect), efficiency indicators (resource and time costs), and robustness indicators (environmental adaptability). While evaluation methods based on these indicators can assess radar anti-jamming effectiveness, they still have several problems: 1) a lack of quantification of radar network synergy effects; 2) the need to estimate interference parameters in practical applications, without considering the impact of parameter estimation errors on the indicators; and 3) how to assign appropriate weights to different indicator parameters. Therefore, considering specific application scenarios, how to construct a more dynamic and targeted evaluation model becomes the research focus of this invention. Summary of the Invention

[0005] The purpose of this invention is to provide a radar anti-jamming decision model evaluation method based on AHC (Analytic Hierarchy Process), which solves the problems of static, single-dimensional and insufficient robustness considerations in the existing evaluation system, and realizes accurate and comprehensive evaluation of radar network anti-jamming strategies.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] An evaluation method for radar anti-jamming decision models based on AHC, characterized by the following steps:

[0008] Step 1: Decompose the overall performance score of the radar anti-jamming decision model, including: effectiveness indicators. Performance indicators With robustness indicators Construct a comprehensive performance score evaluation model ;

[0009] Step 2: Develop performance indicators Efficiency indicators Robustness indicators The sub-indicators are decomposed twice to construct a weighted calculation model for the subordinate sub-indicators; performance indicators Including: SINR increase Interference suppression ratio Signal fidelity Target detection probability Efficiency indicators Includes: policy generation delay Convergence speed Training stability Robustness indicators Including: Node fault tolerance Interference type adaptability ;

[0010] Step 3: Calculate the performance index based on the Analytic Hierarchy Process (AHP). Efficiency indicators Robustness indicators The static weights of each subsidiary sub-indicator in the weighted calculation model;

[0011] Step 4: Classify the interference intensity based on the previous value of the interference suppression ratio, including: weak interference, medium interference, and strong interference. For each interference intensity, construct a sample set and score the overall performance. The training aims to maximize the fit between the actual performance of radar anti-jamming and the target performance, thereby obtaining the effectiveness index. Efficiency indicators Robustness indicators The optimal weighting coefficients are used to obtain a complete comprehensive performance score evaluation model;

[0012] Step 5: Determine the interference intensity level based on the previous value of the interference suppression ratio detected by the radar in real time, match the corresponding comprehensive performance score evaluation model, and calculate the comprehensive performance score.

[0013] Furthermore, in step 1, the comprehensive performance score evaluation model... Represented as:

[0014] ,

[0015] in, , , Performance indicators Efficiency indicators Robustness indicators The weighting coefficients.

[0016] Furthermore, in step 2, stability is trained. Specifically, it is expressed as follows:

[0017] ,

[0018] in, This is the reward value for the Nth round of training. The standard deviation of the reward value. This represents the average reward value.

[0019] Furthermore, in step 2, interference type adaptability Specifically, it is expressed as follows:

[0020] ,

[0021] in, This represents the overall anti-jamming performance score of the radar network under four typical jamming types, representing the radar anti-jamming algorithm model. The standard deviation of the radar network's overall anti-jamming performance score;

[0022] Furthermore, the comprehensive anti-jamming performance score of the radar network is specifically expressed as follows:

[0023] ,

[0024] in, , , These are the normalized values ​​of interference suppression ratio, target detection probability, and SINR improvement, respectively.

[0025] Furthermore, the specific process of step 3 is as follows:

[0026] Regarding performance indicators Efficiency indicators Robustness indicators For any one of them, define the set of legal scales using the 1-9 scaling method, and generate an initial judgment matrix. The weight vector is calculated based on the initial judgment matrix. , The representative indicator types are as follows: ;

[0027] Calculate the largest eigenvalue : , The dimension of the initial judgment matrix;

[0028] Set a consistency threshold: , Indicates the random consistency index;

[0029] Calculate the consistency index : ,when At that time, scale optimization iteration is performed to calculate the pair The scale element with the greatest impact is adjusted in its neighborhood, while the other scales remain unchanged, until... This yields the static weights of each subsidiary sub-indicator in the weighted calculation model.

[0030] Furthermore, in step 4, weak interference is defined as JSR_before < 10dB, medium interference is defined as 10dB ≤ JSR_before ≤ 25dB, and strong interference is defined as JSR_before > 25dB. JSR_before represents the previous value of the interference suppression ratio.

[0031] Furthermore, in step 4, radar is used to detect the distance. Target recognition accuracy With interference suppression effect Constructing radar anti-jamming performance labels for:

[0032] ,

[0033] in, Radar detection range Target recognition accuracy With interference suppression effect The normalization result, They are respectively The preset weighting coefficients;

[0034] Based on actual radar anti-jamming performance labels Quadruple Sample Data Calculate the goodness of linear fit :

[0035] ,

[0036] in, The number of samples in the sample set. For the first The actual performance label of each sample For the first The overall score of each sample This represents the average value of the actual performance labels;

[0037] by To achieve the objective, a genetic algorithm is used to calculate the weights, resulting in an efficiency index. Efficiency indicators Robustness indicators The weight.

[0038] Based on the above technical solution, the beneficial effect of the present invention is that it provides a radar anti-jamming decision model evaluation method based on AHC (Analytic Hierarchy Process), which has the following advantages:

[0039] 1) This invention constructs a three-level structure of target layer-criteria layer-indicator layer, covering the three core dimensions of effectiveness, efficiency and robustness, and adds indicators such as interference type adaptability and training stability, which makes up for the shortcomings of the single dimension of the existing evaluation system and realizes a comprehensive evaluation of anti-interference strategy.

[0040] 2) Combining the AHP method with a dynamic weight adjustment mechanism, the weights of each criterion layer are adaptively adjusted according to the interference intensity, so that the evaluation system can adapt to different dynamic interference scenarios and improve the pertinence and accuracy of the evaluation results.

[0041] 3) The heatmap of indicator contribution enables the interpretability of the evaluation results, clarifies the advantages and disadvantages of the strategy, and provides precise guidance for strategy optimization. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating the radar anti-jamming decision model evaluation method based on AHC in this invention.

[0043] Figure 2 This is a schematic diagram illustrating the static weight calculation process of the auxiliary sub-indices in the radar anti-jamming decision model evaluation method based on AHC in this invention.

[0044] Figure 3 This is a schematic diagram illustrating the dynamic weight adaptive adjustment principle of the radar anti-jamming decision model evaluation method based on AHC in this invention.

[0045] Figure 4 This is a schematic diagram of the MADDPG algorithm network structure in an embodiment of the present invention.

[0046] Figure 5 This is a schematic diagram of the MAAC algorithm network structure in an embodiment of the present invention.

[0047] Figure 6 This is a schematic diagram of the MAPPO algorithm network structure in an embodiment of the present invention.

[0048] Figure 7 This is a schematic diagram of the QMIX algorithm network in an embodiment of the present invention.

[0049] Figure 8 This is a heatmap showing the contribution of indicators to the radar anti-jamming decision model evaluation method based on AHC in this embodiment of the invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0051] This invention provides a radar anti-jamming decision model evaluation method based on AHC (Analytic Hierarchy Process), the process of which is as follows: Figure 1 As shown, the main steps include:

[0052] Step 1: Calculate the overall performance score of the radar anti-jamming decision model. Break it down, including: performance indicators ( ), performance indicators ( ) and robustness index ( ), and then construct an evaluation model, specifically expressed as:

[0053] ,

[0054] in, , , Performance indicators Efficiency indicators Robustness indicators The weighting coefficients; the effectiveness index represents the core effect of anti-jamming; the efficiency index is used to reflect engineering practicality and serves as a constraint condition under the premise of excellent core anti-jamming effect; the robustness index is used to reflect the adaptability of the radar anti-jamming decision model to the environment.

[0055] Step 2: Construct performance indicators Efficiency indicators Robustness indicators Sub-indicators, with clearly defined quantification methods;

[0056] Performance indicators include: SINR increase (ΔSINR) Interference rejection ratio (JSR) Signal fidelity The target detection probability (Pd) is quantified by the weighted average of its four subsidiary indicators, and is expressed as follows:

[0057] ,

[0058] in, These are the weight values ​​of each subsidiary sub-indicator under the performance indicator;

[0059] efficiency indicators include: Strategy generation delay, Convergence speed Training stability, quantified by the weighted average of these three subsidiary metrics, is expressed as follows:

[0060] ,

[0061] in, These are the weight values ​​of each subsidiary sub-indicator under the efficiency index;

[0062] It is particularly important to note that existing efficiency metrics typically only consider convergence speed. However, in practical applications, training stability also needs to be considered. Some algorithms converge quickly, but their rewards fluctuate significantly during training, which can lead to performance instability in real-world engineering deployments. Conversely, some algorithms converge more slowly but exhibit smaller fluctuations, making them more reliable overall. Therefore, this invention sets the coefficient of variation of the reward function during training. As a measure of training stability, that is Specifically, it is expressed as:

[0063] ,

[0064] in, Let the standard deviation of the reward function be . This is the average value. Let be the reward value for the Nth training round; the coefficient of variation of the reward function ranges from [0,1]. The training stability index is negatively correlated with the coefficient of variation of the reward function. This is achieved by constructing... The form of the index makes the index value larger when the stability is higher, and the value range is [0,1]. This embodiment introduces the training stability index to more comprehensively evaluate the algorithm performance.

[0065] Robustness index include: Node fault tolerance Interference type adaptability is quantified by a weighted average of these two subsidiary indicators, and is expressed as follows:

[0066] ,

[0067] in, These are the weight values ​​of each subsidiary sub-indicator under the robustness index;

[0068] It is particularly important to note that existing robustness models typically only consider node fault tolerance, but the core robustness of radar network anti-jamming lies in its adaptability to different types of interference, such as suppression jamming, deception jamming, and frequency sweeping jamming. An excellent decision-making algorithm model should be able to maintain system performance stability under multiple interference types, rather than being applicable only to a single type of interference. Based on this, this invention introduces the reciprocal of the standard deviation of the AHP comprehensive score. As an indicator of interference type adaptability, it is specifically expressed as follows:

[0069] ,

[0070] in, The standard deviation is the reciprocal, and the higher the fit, the larger the value.

[0071] The jamming type adaptability index covers common radar jamming types: suppression (power suppression) jamming, deception (false target) jamming, frequency sweeping jamming, and hybrid (multiple jamming superposition) jamming. This represents the reciprocal of the standard deviation of the algorithm's overall AHP score under the four types of interference; This represents the comprehensive anti-jamming performance score of the radar network under four typical jamming types. It is a quantitative evaluation value of the model's anti-jamming capability under a single jamming type, with a value range of [0,1]. A higher value indicates better anti-jamming performance under that jamming type. Specifically, it is expressed as follows:

[0072] ,

[0073] in, , , These are the normalized values ​​of interference suppression ratio, target detection probability, and SINR improvement, respectively, and their weighting coefficients are fixed values ​​based on criteria in the radar field.

[0074] Step 3: Determine based on the Analytic Hierarchy Process (AHP) Static weights of subsidiary indicators;

[0075] The judgment matrix uses a 1-9 scale, where 1 is equally important, 3 is slightly important, 5 is significantly important, 7 is very important, and 9 is extremely important. The reciprocal represents the opposite importance.

[0076] like Figure 2 As shown, Indicator layer: ΔSINR ( ) and P_d ( Slightly more important than JSR ( ) and fidelity ( (Scale 3), ΔSINR is as important as P_d (Scale 1), JSR is as important as fidelity (Scale 1), then the judgment matrix is ​​represented as:

[0077]

[0078] Indicator layer: Delay ( ), convergence speed ( )and If they are equally important (scale 1), then the judgment matrix is ​​represented as follows:

[0079]

[0080] Metrics layer: Node fault tolerance ( ) and interference adaptability ( If all three are equally important (scale 1), then the judgment matrix is ​​represented as:

[0081]

[0082] By calculating the largest eigenvalue of the judgment matrix ( The system uses eigenvectors and eigenvalues ​​to quantify the relative importance of each indicator, obtains normalized weights, and then performs a consistency check.

[0083] Taking the calculation of the performance indicator layer judgment matrix as an example, the performance indicator layer matrix is ​​as follows:

[0084] ,

[0085] After column normalization, the normalized matrix is ​​obtained:

[0086] ,

[0087] The weight vector is obtained by summing each column of the normalized matrix and then comparing the sum with the number of rows. :

[0088] ,

[0089] Similarly, we get:

[0090] , ;

[0091] For a matrix to have perfect consistency, it essentially means that the relative importance of the indicators must be completely self-consistent, without any conflicts or contradictions. The key to quantifying this consistency is by calculating the largest eigenvalue. Used for consistency testing:

[0092] ,

[0093] in, Representative indicator layer types include: ;for The necessary and sufficient condition for the perfect consistency of the judgment matrix is: , and The smaller the difference, the better the consistency; the larger the difference, the more serious the logical contradiction. hour, A value of 4.2 is better than 4.5. If perfect consistency is not achieved, then further adjustments are needed. ( To determine whether the weight is acceptable:

[0094] If matrix This indicates that matrix consistency is significantly better than that of random matrices, and logical contradictions are acceptable; among them, the consistency index : , It reflects the difference between the current matrix and perfect consistency, but it cannot determine whether this difference is within an acceptable range. It is necessary to use the RI-random consistency index to solve the fairness problem of consistency testing.

[0095] The essence is the average of the randomly generated judgment matrix. The value serves as a baseline. The generation process involves obtaining the average value through a large number of randomized trials using Monte Carlo simulation. A unified RI table has now been established in academia.

[0096]

[0097] Dangruo This indicates that the matrix logic contradictions are too severe. In this case, the algorithm will automatically correct the scale and recalculate consistency until consistency is satisfied, i.e., when... The algorithm is based on The minimization-based intelligent iterative optimization algorithm, through the logic of the scaling neighborhood algorithm and consistency iterative verification, automatically adjusts the scale value of the judgment matrix within the rules of the 1-9 scaling method (1-9 and its reciprocal, representing the relative importance between indicators) until the condition is met. Consistency requirements:

[0098] When the initial judgment matrix Determine its order Obtain the corresponding table Value, set consistency threshold: ,

[0099] And define the scaled valid set This corresponds to the rule range of the 1-9 scaling method; all scaling adjustments are made only from... Take the value from, when At this point, the algorithm enters the scaling optimization iteration, by calculating the... The most influential scaling element To avoid blind optimization, Perform neighborhood adjustment, keeping the other scales unchanged, and set the neighborhood adjustment step size to 1, i.e., within the valid set of scales. Only adjacent values ​​are adjusted until the output is reached. ;

[0100] Finally obtained The weights of the subordinate sub-indicators;

[0101] Step 4: Classify the interference intensity based on the previous interference suppression ratio (JSR_before), including: weak interference (JSR_before < 10dB), medium interference (10dB ≤ JSR_before ≤ 25dB), and strong interference (JSR_before > 25dB). Construct a sample set of interference intensity-indicator-performance for different interference intensities to obtain a comprehensive score. With the goal of maximizing the fit with the actual performance of radar anti-jamming, the algorithm trains and learns on the sample set through optimization to autonomously generate the optimal weight coefficients corresponding to each jamming intensity level. Each weight coefficient satisfies Establish a dynamic weight library;

[0102] The core of the autonomous generation of dynamic weight coefficients based on interference intensity is to replace manual weight setting with a data-driven approach. This involves training an algorithmic model on measured radar anti-jamming data under different interference intensities to establish a dynamic weight library that can be matched in real time. The specific steps are as follows:

[0103] First, the algorithm adopts a hierarchical training strategy: the samples are divided into three independent sample sets according to the interference intensity, and each sample set is optimized and solved separately to generate weight coefficients specific to each intensity, rather than a single global weight, so as to ensure accurate matching between the weights and the scene.

[0104] Secondly, in addition to the three types of indicators and three data points In addition, a new actual performance label has been added. (Range detected by radar) Target recognition accuracy Actual effect of interference suppression The comprehensive quantitative value obtained by normalizing the three measured indicators: In particular, considering the requirements of radar in actual combat—first suppressing, then jamming, then ensuring detection, and finally accurately identifying—[the following is likely a separate point:] The values ​​are 0.25, 0.25, and 0.5 respectively. These are the normalized values ​​of the three indicators, forming quadruplet data. Quadruplet data are then collected for different interference intensities. ;

[0105] Then, the typical MARL algorithm was selected as the experimental platform, and experiments were conducted for different intensities, using the most commonly used linear fit goodness-of-fit metric. :

[0106] ,

[0107] As the core objective function, where, The number of samples in the sample set. For the first The actual performance label of each sample For the first The overall score of each sample Actual performance label The average of the sums;

[0108] Finally, based on the existing random search algorithm for natural selection: the genetic algorithm, the weights are calculated to achieve the desired result. That is to let Get as close to 1 as possible to achieve the best result. Weight values, such as Figure 3 As shown;

[0109] Step 5: Determine the interference intensity level based on the JSR_before detected by the radar in real time, and determine the corresponding matching. The weight values ​​are used to calculate the score S. In addition, the contribution of each indicator to the overall score can be calculated: indicator contribution = indicator normalized value × AHP global weight. The contribution ratio of each indicator is visualized through a heatmap, which intuitively shows the advantages and disadvantages of the strategy.

[0110] The beneficial effects of the present invention will be explained in detail below with reference to simulation tests.

[0111] Experiments were conducted to validate different decision model algorithms ('MADDPG', 'MAPPO', 'QMIX', 'MAAC') under MARL. The different algorithms were evaluated by combining the evaluation system scores with the indicator effect graphs, thus verifying the scientific validity of the evaluation system.

[0112] The core idea of ​​MADDPG (Multi-Agent Deep Deterministic Policy Gradient) is based on centralized training and distributed execution (CTDE). The critic network is trained using global information, while the actor network makes decisions using only local observations. Figure 4 As shown; in an actor network, each agent... Based on local observations Output Action To implement a deterministic strategy; in a critic network, a centralized critic receives observations from all agents. and actions Calculate the Q value During training, the actor is updated using the gradient of the critic's Q-value (maximizing) The critic is updated via TD error; during execution, only the actor network is used, and decisions are made independently.

[0113] The core idea of ​​MAAC (Multi-Agent Advantage Actor-Critic) is to extend A2C to multi-agent scenarios by introducing advantage functions and credit allocation mechanisms, such as... Figure 5 As shown; in an actor network, each agent... Based on local observations Output Action In a critic network, the global state value is calculated. or centralized Value, through The advantage function is used to distinguish individual contributions; during training, the actor utilizes its own advantage. By maximizing the policy gradient, the critic updates the value estimate through the reward error, and the advantage function helps the agent distinguish the impact of its own actions on the global situation.

[0114] The core idea of ​​MAPPO (Multi-Agent Proximal Policy Optimization) is a multi-agent version of PPO, which uses gradient stabilization training through policy clipping and supports centralized value evaluation, such as... Figure 6 As shown; in an actor network, each agent... Output action distribution The old strategy is retained for calculating the clip ratio; in the critic network, the centralized critic is based on global information (state). (or joint observation) Calculation value Used to calculate advantage Training, on the other hand, involves editing the target... The actor and critic are updated using MSE loss; the stability of PPO makes it perform better in complex multi-agent scenarios.

[0115] The core idea of ​​QMIX (Q-Mixing Network) is a value decomposition method that decomposes the global Q-value into monotonic combinations of individual Q-values ​​to solve the credit allocation problem, such as... Figure 7 As shown; in an individual Q-network, each agent Based on local observation and action Output individual Q value In a mixing network, the Q-values ​​of all individuals are... With global state Combined, output the global Q value. It also ensures the monotonicity of the decomposition, and the global Q-value increases as the individual Q-value increases. During training, the hybrid network and the individual Q-network are updated through TD error. During execution, each agent makes greedy decisions based on its individual Q-value, eliminating the need for a centralized actor and enabling purely distributed execution.

[0116] In this embodiment, the model is trained using radar simulation signals and datasets of three different types of interference (narrowband interference, broadband interference, and hybrid interference). On different MARL algorithms, the AHC evaluation index system is used to evaluate and quantify the different algorithms. To assess the merits of the algorithms, the same index system is used to evaluate the radar network anti-interference capabilities of the algorithm model. For the output results, the contribution ratio of each index is visualized through a heatmap, which intuitively shows the advantages and disadvantages of the strategy.

[0117] Specifically, the simulation test conditions are as follows: The algorithm model in this embodiment starts from the electronic countermeasures of radar, generates three types of interference signals to interfere with radar signals, and uses the anti-interference capability of the algorithm to output anti-interference decisions. The anti-interference effect of different algorithms is quantitatively evaluated using this invention. The invention is applied to the basic model and compared with traditional indicators to highlight the engineering practicality and innovation of this invention.

[0118] Table 1 shows the differences of the four algorithms in terms of P_D, average latency, fault tolerance, training stability, and interference adaptability. As can be seen from the table, without the comprehensive AHP scoring system, each indicator has its own advantages. Some have lower latency, some have better training stability, and some have better adaptability, making it impossible to intuitively show which algorithm is better. The evaluation index system using the AHC hierarchical analysis method is significantly better than the traditional method of simply looking at the indicators. This indicates that when the model's AHP score is high, this type of algorithm is more suitable for the current application scenario, and the model has a greater advantage.

[0119] Table 1

[0120]

[0121] The traditional and representative evaluation method for MARL algorithms is the EpyMARL method. EpyMARL is an open-source framework designed for multi-agent reinforcement learning (MARL). Its evaluation system is not based on the calculation of a single index, but rather on a multi-dimensional evaluation system built around the characteristics of multi-agent cooperation, algorithm training dynamics, and environmental interaction effects. The results of evaluating the above four algorithms using the EpyMARL method are shown in Table 2.

[0122] Table 2

[0123]

[0124] As shown in Table 2, among the four algorithms, MAPPO, MADDPG, and MAAC have similar scores, with MADDPG showing better stability. From the perspective of users employing radar anti-jamming strategies, there is no intuitive bias in the selection of the algorithm. In addition, the EpyMARL method does not specifically evaluate radar anti-jamming related indicators.

[0125] In summary, this invention, with multi-objective decision-making and dynamic scenario adaptation at its core, is more suitable for evaluation scenarios that require balancing complex needs, subjective weights, and scenario priorities. Its dynamic weighting mechanism, which better aligns with real-world scenarios and adapts to complex needs and priorities, is the core advantage of this invention and the fundamental difference between it and EpyMARL. Furthermore, multi-agent evaluation often suffers from conflicting metrics; for example, improving cooperation may lead to increased communication latency. The hierarchical structure of this invention effectively resolves such conflicts. This invention not only calculates metrics but also transforms the engineering requirements for radar anti-jamming into a quantifiable, interpretable, and decision-making evaluation system through a hierarchical structure, dynamic weights, and multi-objective trade-offs.

[0126] like Figure 8 As shown in the figure, the evaluation index system of this invention outputs a graph showing the contribution of different algorithm models to their performance. As can be seen from the graph, the darker the color, the higher the contribution of this type of index in the evaluation system of this invention. This indicates to a certain extent that the index is an advantage of this type of algorithm. If you focus on this index, you can intuitively see the effect of comparing it with other algorithms.

[0127] The above description is merely a specific embodiment of the present invention. Any feature disclosed in this specification may be replaced by other equivalent or similar features unless otherwise specified. All disclosed features, or steps in all methods or processes, may be combined in any way except for mutually exclusive features and / or steps.

Claims

1. A radar anti-jamming decision model evaluation method based on AHC, characterized in that, Includes the following steps: Step 1: Decompose the overall performance score of the radar anti-jamming decision model, including: effectiveness indicators. Performance indicators With robustness indicators Construct a comprehensive performance score evaluation model ; Step 2: Develop performance indicators Efficiency indicators Robustness indicators The sub-indicators are decomposed twice to construct a weighted calculation model for the subordinate sub-indicators; performance indicators Including: SINR increase Interference suppression ratio Signal fidelity Target detection probability Efficiency indicators Includes: policy generation delay Convergence speed Training stability Robustness indicators Including: Node fault tolerance Interference type adaptability ; Step 3: Calculate the performance index based on the Analytic Hierarchy Process (AHP). Efficiency indicators Robustness indicators The static weights of each subsidiary sub-indicator in the weighted calculation model; Step 4: Classify the interference intensity based on the previous value of the interference suppression ratio, including: weak interference, medium interference, and strong interference. For each interference intensity, construct a sample set and score the overall performance. The training aims to maximize the fit between the actual performance of radar anti-jamming and the target performance, thereby obtaining the effectiveness index. Efficiency indicators Robustness indicators The optimal weighting coefficients are used to obtain a complete comprehensive performance score evaluation model; Step 5: Determine the interference intensity level based on the previous value of the interference suppression ratio detected by the radar in real time, match the corresponding comprehensive performance score evaluation model, and calculate the comprehensive performance score.

2. The radar anti-jamming decision model evaluation method based on AHC according to claim 1, characterized in that, In step 1, the comprehensive performance score evaluation model Represented as: ， in, , , Performance indicators Efficiency indicators Robustness indicators The weighting coefficients.

3. The radar anti-jamming decision model evaluation method based on AHC according to claim 1, characterized in that, In step 2, training stability Specifically, it is expressed as follows: ， in, This is the reward value for the Nth round of training. The standard deviation of the reward value. This represents the average reward value.

4. The radar anti-jamming decision model evaluation method based on AHC according to claim 1, characterized in that, In step 2, interference type adaptability Specifically, it is expressed as follows: ， in, This represents the overall anti-jamming performance score of the radar network under four typical jamming types, representing the radar anti-jamming algorithm model. The standard deviation of the radar network's overall anti-jamming performance score.

5. The radar anti-jamming decision model evaluation method based on AHC according to claim 4, characterized in that, The radar network's overall anti-jamming performance score is specifically expressed as follows: ， in, , , These are the normalized values ​​of interference suppression ratio, target detection probability, and SINR improvement, respectively.

6. The radar anti-jamming decision model evaluation method based on AHC according to claim 1, characterized in that, The specific process of step 3 is as follows: Regarding performance indicators Efficiency indicators Robustness indicators For any one of them, define the set of legal scales using the 1-9 scaling method, and generate an initial judgment matrix. ; The weight vector is calculated based on the initial judgment matrix. , The representative indicator types are as follows: ; Calculate the largest eigenvalue : , The dimension of the initial judgment matrix; Set a consistency threshold: , Indicates the random consistency index; Calculate the consistency index : ,when At that time, scale optimization iteration is performed to calculate the pair The scale element with the greatest impact is adjusted in its neighborhood, while the other scales remain unchanged, until... This yields the static weights of each subsidiary sub-indicator in the weighted calculation model.

7. The radar anti-jamming decision model evaluation method based on AHC according to claim 1, characterized in that, In step 4, weak interference is defined as JSR_before < 10dB, medium interference is defined as 10dB ≤ JSR_before ≤ 25dB, and strong interference is defined as JSR_before > 25dB. JSR_before represents the previous value of the interference suppression ratio.

8. The radar anti-jamming decision model evaluation method based on AHC according to claim 1, characterized in that, In step 4, radar is used to detect the distance. Target recognition accuracy With interference suppression effect Constructing radar anti-jamming performance labels for: ， in, Radar detection range Target recognition accuracy With interference suppression effect The normalization result, They are respectively The preset weighting coefficients; Based on actual radar anti-jamming performance labels Quadruple Sample Data Calculate the goodness of linear fit : ， in, The number of samples in the sample set. For the first The actual performance label of each sample For the first The overall score of each sample This represents the average value of the actual performance labels; by To achieve the objective, a genetic algorithm is used to calculate the weights, resulting in an efficiency index. Efficiency indicators Robustness indicators The weight.

Images