Key event detection-based wargaming process key maneuver mining method and device

Abstract

The application discloses a kind of based on key event detection's wargaming process key battle method mining method and device, comprising: the multi-resolution situation feature abstraction of the wargaming data collected, obtain the battlefield situation feature set of combat unit under uniform dimension;Known key event in wargaming data is extracted based on battlefield situation feature set using weak supervision self-training method, and specific key event set is obtained;Based on battlefield situation feature set and specific key event set, new key event in wargaming data is extracted using new key event identification method based on deep hidden layer feature, and new key event set is obtained by labeling new key event;Each key event in specific key event set and new key event set is combined according to time sequence relationship, forms multiple battle method set, completes key battle method mining.The present application improves the timeliness and accuracy of key battle method mining.

Description

Technical Field

[0001] This invention belongs to the field of large-scale campaign-level wargaming and intelligent game technology, specifically involving a method and device for mining key tactics in wargaming processes based on key event detection. Background Technology

[0002] Using wargaming simulation for combat exercises is a scientific and effective method for designing future wars and forging combat readiness. It plays a significant role in large-scale joint military exercises and training, daily command training for commanders, and research on combat theory and tactics.

[0003] As the pace and complexity of modern warfare continue to accelerate, human decision-making is increasingly inadequate to keep pace with the rapidly changing battlefield landscape. Future warfare urgently requires rapid, automated, and autonomous decision-making, necessitating the extension of intelligent technologies to the human brain to enhance command and information system capabilities and adapt to the high-speed, complex, and ever-changing battlefield environment. In recent years, the rapid development of artificial intelligence (AI) technology and the introduction of cutting-edge technologies such as deep learning and reinforcement learning have, to some extent, improved the automation and intelligence of combat simulations. Some of the tactics emerging from AI-powered decision-making agents during simulations can provide valuable insights for commanders' command and control decisions.

[0004] Because wargaming simulations are often large-scale and lengthy, involving a vast number and diverse types of combat entities and weapon platforms, it is extremely difficult for commanders to pinpoint key actions from the high-frequency, long-term decisions of AI decision-making agents. This is both time-consuming and prone to inaccurate identification of key tactics. Therefore, there is an urgent need for an innovative tactic mining method that can automatically and intelligently extract key battlefield situations and critical actions taken by commanders before and after specific points from the simulated combat data after the simulation. This would deepen commanders' understanding of the game-theoretic decision-making made by AI algorithms, improve the efficiency of commanders' command training using wargaming, and help them discover new tactics to enhance their command skills. Summary of the Invention

[0005] One of the objectives of this invention is to provide a method for mining key tactics in wargaming simulations based on key event detection, which is used to address the need for strategy abstraction in large-scale intelligent games and solve the problems of poor interpretability and unintuitive decision-making processes in intelligent decision-making schemes generated by reinforcement learning in typical combat scenarios.

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

[0007] A method for mining key tactics in wargaming simulations based on key event detection, comprising:

[0008] Multi-resolution situational feature abstraction is performed on the collected wargaming data to obtain a battlefield situational feature set of combat units under a unified dimension;

[0009] Based on the battlefield situation feature set, a weakly supervised self-training method is used to extract known key events from the wargaming data and obtain a specific key event set.

[0010] Based on the battlefield situation feature set and specific key event set, a novel key event identification method based on deep hidden layer features is used to extract new key events from war game simulation data, and the new key events are labeled to obtain a new key event set.

[0011] By combining key events from specific key event sets and new key event sets according to their chronological relationships, a multi-faceted set of tactics is formed, thus completing the discovery of key tactics.

[0012] Several alternative methods are provided below, but they are not intended as additional limitations on the overall solution above. They are merely further additions or optimizations. Provided there are no technical or logical contradictions, each alternative method can be combined individually with respect to the overall solution above, or multiple alternative methods can be combined with each other.

[0013] Preferably, the process of abstracting multi-resolution situational features from the collected wargame data includes:

[0014] Collaborative features are obtained by processing the original battlefield coordination information in wargame data using dummy coding, interval scaling, and discretization methods.

[0015] The static features of combat units in the collaborative features are augmented using dimensionality-enhancing methods, and the dynamic features of combat units in the collaborative features are reduced using dimensionality-reduction methods.

[0016] The final battlefield situation feature set is obtained by processing static and dynamic features based on the feature selection method. The battlefield situation feature set includes a situation-command feature sequence arranged according to the decision time order.

[0017] Preferably, the process of processing the original battlefield coordination information in the wargaming data using dummy coding, interval scaling, and discretization methods to obtain coordination features includes:

[0018] Dumb encoding of features is performed on the original battlefield coordination information in the wargaming data to obtain the coordination features of combat units when carrying out combat missions.

[0019] The dimensionless method of interval scaling is used to process the collaborative features, transforming the differentiated format data of different combat units into unified format data;

[0020] Discretize the interval-scaled collaborative features to obtain the final collaborative features.

[0021] Preferably, the method of increasing the dimensionality of the static features of combat units in the collaborative features and reducing the dimensionality of the dynamic features of combat units in the collaborative features includes:

[0022] The static characteristics of combat units are augmented using polynomial expansion and kernel functions.

[0023] Principal component analysis and linear discriminant analysis are used to reduce the dimensionality of dynamic data of combat units.

[0024] Preferably, the method of extracting known key events from wargaming data using a weakly supervised self-training method based on the battlefield situation feature set, and obtaining a specific set of key events, includes:

[0025] Clustering is performed on data with concentrated battlefield situation characteristics to obtain clustering results;

[0026] Based on the spatial distribution of data represented by the clustering results, a semi-supervised support vector machine classifier is trained in combination with a labeled combat key event dataset as a classification and recognition model, and the classification and recognition model is used to label the data in the battlefield situation feature set.

[0027] Based on prior trigger flags, data whose labels match the prior trigger flags are selected from the already labeled battlefield situation feature set to form a specific set of key events.

[0028] Preferably, the extraction of novel key events from wargaming data using a novel key event identification method based on deep hidden layer features, based on a battlefield situation feature set and a specific key event set, includes:

[0029] The LSTM network is trained using a specific set of key events. The feature extraction layer in the trained LSTM network is then combined with a multi-head attention network to obtain a hidden layer feature extraction model.

[0030] A linear classification model is constructed based on the definition of the new key event;

[0031] The battlefield situation features from the battlefield situation feature set are input into the hidden layer feature extraction model to obtain the hidden layer features of the battlefield situation features.

[0032] The extracted hidden features and the known hidden features of key events are input into the linear classification model, and new key events in the battlefield situation feature set are determined based on the output of the linear classification model.

[0033] Preferably, the construction of a linear classification model based on the definition of novel key events includes:

[0034] If the similarity between the hidden feature to be judged and the hidden feature of the known key event meets the first threshold, and the action decision corresponding to the hidden feature to be judged is different from that of the hidden feature of the known key event, and the reward obtained by the hidden feature to be judged is higher than the second threshold, then the hidden feature to be judged is a new type of key event.

[0035] Alternatively, if the difference between the hidden feature to be judged and the hidden feature of the known key event meets the third threshold, and the action decision corresponding to the hidden feature to be judged and the hidden feature of the known key event is the same, and the reward obtained by the hidden feature to be judged is higher than the fourth threshold, then the hidden feature to be judged is a new type of key event.

[0036] Preferably, the step of annotating the novel key events to obtain a novel key event set includes:

[0037] The elements that constitute battlefield events during a limited wargaming simulation;

[0038] Based on the battlefield element weights contained in the new key events output by the multi-head attention network, the core elements are determined according to the battlefield element weights, and discrete values ​​are determined for the core elements, resulting in multiple situation description keywords composed of core elements and discrete values.

[0039] Based on multiple situation description keywords, a keyword template concatenation method is used to obtain the annotation information of new key events, and the annotated new key events are used as a new set of key events.

[0040] The present invention provides a method for mining key tactics in wargaming simulations based on key event detection. This method automatically and intelligently extracts key battlefield situations and key actions taken by commanders before and after the simulation combat data after the simulation ends. This deepens the understanding of commanders and combat personnel regarding the game-theoretic decision-making made by artificial intelligence algorithms, improves the efficiency of commanders using wargaming simulations for command training, and helps them discover new tactics to enhance their command skills.

[0041] The second objective of this invention is to provide a key tactical mining device based on key event detection for wargame simulations, which addresses the need for abstracting strategies in large-scale intelligent games and solves the problems of poor interpretability and unintuitive decision-making processes in typical combat scenarios when intelligent decision-making schemes generated through reinforcement learning for game adversarial purposes are not readily interpretable.

[0042] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0043] A device for mining key tactics in wargaming simulations based on key event detection includes a processor and a memory storing a number of computer instructions. When the computer instructions are executed by the processor, they implement the steps of the method for mining key tactics in wargaming simulations based on key event detection. Attached Figure Description

[0044] Figure 1 This is a flowchart of the key tactics mining method in wargaming simulation based on key event detection in this invention.

[0045] Figure 2 This is an example diagram of dimensionality reduction using linear discriminant analysis applied in this invention;

[0046] Figure 3 This is a flowchart of the key event extraction process based on weakly supervised self-training in this invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0049] To better assist commanders in analyzing and understanding large-scale wargaming simulation data, and to improve their daily training efficiency and wartime decision-making capabilities, this embodiment proposes a key tactics mining method based on key event detection in the wargaming simulation process. This provides a research and solution to the challenge of understanding agent behavior in large-scale wargaming simulations, and improves the accuracy of key tactics mining.

[0050] For ease of understanding, this embodiment uses a military chess game simulation as an example for illustration. Figure 1 As shown in the figure, the implementation steps of the key tactics mining method based on key event detection in the wargaming process in this embodiment are as follows:

[0051] Step 1: Perform multi-resolution situational feature abstraction on the collected wargame simulation data to obtain a battlefield situational feature set of combat units under a unified dimension.

[0052] Large-scale military chess simulation data suffers from various problems, including massive information volume, high redundancy, and inconsistent data formats. This embodiment addresses these issues by applying dummy coding, interval scaling, and discretization to process the tactical coordination characteristics. It also employs polynomial expansion, kernel function dimensionality enhancement, and PCA / LDA dimensionality reduction methods to encode and process multidimensional static features such as unit name, type, and performance parameters, as well as dynamic features such as latitude, longitude, orientation, and speed. Furthermore, it utilizes feature selection methods like Filter, Wrapper, and Embedded to perform multi-resolution situational feature engineering, mitigating the impact of special mechanisms, simulation mechanisms, and effect delays in campaign-level adversarial scenarios, thereby achieving effective abstraction of crucial situational information.

[0053] 1) Process the tactical coordination characteristics.

[0054] For large-scale combat simulation scenarios, the first step is to abstract the characteristics of multi-service coordination:

[0055] ① In order to transform qualitative multi-service coordination data into quantitative features, dummy coding of the original battlefield coordination information is performed to obtain the coded features (coordination features) of different combat units such as different branches of service and troops when carrying out specific combat missions.

[0056] ② Since the collaborative feature formats differ among different combat units, a dimensionless method of interval scaling is used to process the collaborative features, so that the differentiated format data of different combat units can be converted into a unified specification.

[0057] ③ In order to improve the expressive power of tactical coordination features and enhance the robustness of the model to abnormal data, the coordination features after interval scaling are discretized and nonlinearity is introduced to solve the problem of information redundancy.

[0058] After a dummy coding process for battlefield situation information, qualitative data on the battlefield is encoded into processable quantitative data. Considering the complexity and diversity of the battlefield environment and the unavoidable occurrence of outlier data, the data is discretized using an equal-frequency discretization method to enhance the robustness of subsequent models to outlier data.

[0059] 2) Dimensionality increase and decrease of features.

[0060] Battlefield situational data is complex and diverse, encompassing static features such as unit names, types, and performance parameters, as well as dynamic features like latitude, longitude, altitude, orientation, and speed. Both static and dynamic features are largely non-linear. When data is non-linear, directly using models like linear regression can significantly lead to underfitting. Furthermore, because static features such as unit names, types, and performance parameters have low dimensionality, to comprehensively consider both static and dynamic characteristics of the battlefield situation, multinomial expansion and kernel functions are used to transform the existing static features of the battlefield situation, mapping them to a higher-dimensional space, thus enabling them to fit a wider range of data.

[0061] Because the dynamic characteristics of combat units, such as latitude, longitude, altitude, orientation, and speed, have high dimensionality, principal component analysis (PCA) and linear discriminant analysis (LDA) are used to reduce the dimensionality of the dynamic characteristics of combat units, so that the static and dynamic characteristics of combat units can be consistently encoded and processed.

[0062] The dynamic characteristics of combat units contain a great deal of redundant data. To extract key battlefield situational information, principal component analysis (PCA) is first used to calculate the covariance matrix of the data matrix composed of the dynamic characteristic data of all combat units. This yields the eigenvalues ​​and eigenvectors of the covariance matrix. The matrix formed by the eigenvectors corresponding to the k features with the largest eigenvalues ​​(i.e., the largest variance) is then selected to transform the dynamic data matrix of the combat units into a new feature space, achieving dimensionality reduction. Furthermore, linear discriminant analysis is used to further reduce the dimensionality of the dynamic characteristics of the combat units after PCA processing. The data is projected onto a lower dimension, aiming for the projection points of each category to be as close as possible, while maximizing the distance between the category centers of different categories. Figure 2 As shown.

[0063] By using polynomial expansion and kernel functions to increase the dimensionality and encode the static features of combat units, and by using methods such as PCA and LDA to reduce the dimensionality of the dynamic features of combat units, we can obtain the battlefield situation features of combat units under a unified dimension. This is beneficial for the abstraction of important battlefield situation information.

[0064] 3) Multi-resolution situational feature engineering is carried out based on feature selection methods such as Filter, Wrapper, and Embedded to solve the impact of special mechanisms, simulation mechanisms and effect delays in campaign-level confrontation scenarios, and to achieve effective abstraction of important situational information.

[0065] Step 2: Based on the battlefield situation feature set, a weakly supervised self-training method is used to extract known key events from the wargaming data and obtain a specific key event set.

[0066] To identify specific key tactics and discover new (novel) key tactics, it is first necessary to use the abstracted battlefield situation characteristics and agent operations to identify and extract key events.

[0067] Based on the digitally abstracted situation-command feature sequence, this study comprehensively utilizes the situation-command features at a single decision moment and the situation-command feature sequence data with sequential form. By applying a weakly supervised self-training algorithm, it analyzes the data generated during large-scale long-term game simulations from the perspectives of the correlation between situation and action, and the correlation between action commands and battlefield changes, thereby locating key events in the simulated battle situation.

[0068] For events in the simulation and deduction process of military chess, this embodiment follows the event definition in natural language processing and consists of two parts: trigger flags and constituent elements. The trigger flag is the element that best represents the occurrence of an event. In this example, it is specifically the situation-instruction feature pair at a certain moment, which is the main feature that determines the event category. The constituent elements are the auxiliary response processes before and after the event trigger flag. Together with the trigger flag, they constitute the entire framework of the event.

[0069] Considering that it is impractical to effectively annotate a large number of events during large-scale military chess simulations, this embodiment employs a weakly supervised method for event detection. Based on existing, relatively small-scale datasets of key combat events built with technical guidance from professional military personnel, as well as a large dataset of military chess simulations, a semi-supervised approach is used to train an event classification and recognition model. For example... Figure 3 As shown, the specific operation is as follows:

[0070] First, constrained clustering is performed on the unlabeled military chess simulation data. In the process of obtaining the clustering results, the correlation between time information, entity information and battlefield environment information is considered at the same time. The specific idea is that in a specific battlefield environment, if the entity states are similar at similar times, they are likely to trigger the same event, and constrained clustering is performed in this way.

[0071] After clustering in this way, a simple semi-supervised support vector machine (SVM) classifier is trained by combining it with a labeled dataset of key combat events. The data spatial distribution of the unlabeled inference data is utilized to ensure that the final classification hyperplane of the semi-supervised SVM classifier aligns with the distribution of the unlabeled data. Based on the classification results, the unlabeled data is labeled, and trigger markers (typical battlefield situation - typical behavioral decisions) are initially selected from the sequence data as the feature key event set, representing the most representative battlefield situation and agent-issued commands in a time-series data segment. This yields complete labeled data for event extraction, which is then jointly trained with the labeled data to improve model performance.

[0072] Step 3: Based on the battlefield situation feature set and specific key event set, a new key event identification method based on deep hidden layer features is used to extract new key events from the war game simulation data, and the new key events are labeled to obtain a new key event set.

[0073] Step 31: Novel Key Event Recognition Based on Deep Hidden Layer Feature Extraction.

[0074] In the previous step, this embodiment used a weakly supervised self-training method to initially classify key events in large-scale simulation data according to prior categories. However, this classification inevitably resulted in some novel key events being categorized into existing categories or being missed. Therefore, this embodiment proposes a novel key event identification method based on deep hidden layer features. Leveraging the efficiency and accuracy of deep recurrent neural networks in extracting features from sequence data, it further detects novel key events from three perspectives: situational features, behavioral features, and reward features. This method also identifies missed key events from the simulation data and further subdivides novel key events from previously located specific key events.

[0075] 1) Using the specific key event set obtained in step 2, retrain the LSTM network that has been pre-trained on a public natural language processing dataset to achieve optimization and correction, and extract the feature extraction layer, i.e. the backbone network, as the main body of subsequent feature analysis.

[0076] 2) A linear classification model is constructed based on the definitions of several novel key events. The core idea of ​​this definition is: if a key event makes a different action decision when its hidden feature similarity to a known key event meets a specified threshold, and the subsequent reward is higher than the specified threshold, then it is a novel key event; if a key event makes the same action decision as one in the existing set of key events when its hidden feature difference meets a specified threshold, and the subsequent reward is higher than the specified threshold, then it is a novel key event. Based on these assumptions, a linear classification model is formed.

[0077] 3) Based on the linear classification model constructed in step 2) and the backbone network trained in step 1) combined with the multi-head attention network, and based on the similarity of hidden layer features, the fuzzy matching algorithm is used to reclassify the inferred data by setting a reasonable threshold, thereby extracting the key events that were not previously extracted and some new key events that were confused with the defined key events. The specific steps are as follows: Input a piece of inferred data, that is, the battlefield situation features (situation-command features) in the battlefield situation feature set, and use the feature extraction network formed by the backbone network combined with the multi-head attention network to obtain the hidden layer features of the inferred data. The hidden layer features and the hidden layer features of the existing known key events are fed into the linear classification model constructed in step 2) to determine whether it is a new type of key event.

[0078] Step 32: Automatic labeling of new key tactics based on intelligent keyword description and concatenation.

[0079] It is impractical to manually label each event fragment extracted in step 31 for the novel key events. This embodiment uses keyword intelligent description and concatenation technology to automatically label them. For the situation-instruction feature sequence divided into novel key events, intelligent description technology is applied, relying on the attention mechanism, to automatically describe the most core enemy combat entity information, combat intention information, our strength information, and our instruction information in the situation using keywords. Based on the keyword concatenation method, the label information of the novel key tactics is obtained.

[0080] 1) Utilize prior knowledge to define the elements that constitute battlefield events, such as the type of target entity, the number of target entities, the location of target entities, the collaborative relationship between target entities, whether the target entity is being detected for the first time, and our current strategic deployment. Then, define the core behavior of the event and limit it to a few sub-behaviors that the commander is familiar with.

[0081] 2) Combining the battlefield element weights output by the multi-head attention network used in sub-step 3) of step 31, determine the core elements that need to be described, and assign discrete values ​​to them to obtain situation description keywords. The elements are the main body, and the discrete values ​​are the descriptions. The combination of the two forms the situation description keywords.

[0082] 3) Apply the keyword template concatenation method to obtain the annotation information of new key events.

[0083] Step 4: Combine the key events in the specific key event set and the new key event set according to their chronological relationship to form a multi-tactic set, and complete the key tactic mining.

[0084] This embodiment automatically and intelligently extracts key battlefield situations and key actions taken by commanders before and after the point of the simulation from the simulated combat data after the simulation ends. This improves the timeliness and accuracy of key tactics discovery, deepens the understanding of commanders and combat personnel on the game-theoretic decision-making made by artificial intelligence algorithms, improves the efficiency of commanders using wargaming for command training, and helps them discover new tactics to enhance their command skills.

[0085] In another embodiment, this application also provides a device for mining key tactics in wargaming based on key event detection, including a processor and a memory storing a plurality of computer instructions, wherein the computer instructions, when executed by the processor, implement the steps of the method for mining key tactics in wargaming based on key event detection.

[0086] For specific limitations on the key tactics mining device in wargaming process based on key event detection, please refer to the limitations on the key tactics mining method in wargaming process based on key event detection mentioned above, which will not be repeated here.

[0087] The memory and processor are electrically connected directly or indirectly to enable data transmission or interaction. For example, these components can be electrically connected to each other via one or more communication buses or signal lines. The memory stores a computer program that can run on the processor, which implements the key tactic mining method for wargaming processes based on key event detection in this embodiment of the invention by running the computer program stored in the memory.

[0088] The memory may be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The memory stores the program, and the processor executes the program upon receiving an execution instruction.

[0089] The processor may be an integrated circuit chip with data processing capabilities. The aforementioned processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor.

[0090] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0091] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for mining key tactics in wargaming simulations based on key event detection, characterized in that, The method for mining key tactics in wargaming simulations based on key event detection includes: Multi-resolution situational feature abstraction is performed on the collected wargaming data to obtain a battlefield situational feature set of combat units under a unified dimension; Based on the battlefield situation feature set, a weakly supervised self-training method is used to extract known key events from the wargaming data and obtain a specific key event set. Based on the battlefield situation feature set and specific key event set, a novel key event identification method based on deep hidden layer features is used to extract new key events from war game simulation data, and the new key events are labeled to obtain a new key event set. By combining key events from specific key event sets and new key event sets according to their chronological relationships, a multi-faceted set of tactics is formed, thus completing the discovery of key tactics. The process of abstracting multi-resolution situational features from the collected wargame simulation data includes: Collaborative features are obtained by processing the original battlefield coordination information in wargame data using dummy coding, interval scaling, and discretization methods. The static features of combat units in the collaborative features are augmented using dimensionality-enhancing methods, and the dynamic features of combat units in the collaborative features are reduced using dimensionality-reduction methods. The final battlefield situation feature set is obtained by processing static and dynamic features based on the feature selection method. The battlefield situation feature set includes a situation-command feature sequence arranged according to the decision time order. Specifically, the method of extracting known key events from wargaming data using a weakly supervised self-training approach based on the battlefield situation feature set, and obtaining a specific set of key events, includes: Clustering is performed on data with concentrated battlefield situation characteristics to obtain clustering results; Based on the spatial distribution of data represented by the clustering results, a semi-supervised support vector machine classifier is trained in combination with a labeled combat key event dataset as a classification and recognition model, and the classification and recognition model is used to label the data in the battlefield situation feature set. Based on prior triggering flags, data with labels matching the prior triggering flags are selected from the already labeled battlefield situation feature set to form a specific set of key events; The extraction of novel key events from wargaming data using a novel key event identification method based on deep hidden layer features, based on a battlefield situation feature set and a specific key event set, includes: The LSTM network is trained using a specific set of key events. The feature extraction layer in the trained LSTM network is then combined with a multi-head attention network to obtain a hidden layer feature extraction model. A linear classification model is constructed based on the definition of the new key event; The battlefield situation features from the battlefield situation feature set are input into the hidden layer feature extraction model to obtain the hidden layer features of the battlefield situation features. The extracted hidden features and the known hidden features of key events are input into the linear classification model, and new key events in the battlefield situation feature set are determined based on the output of the linear classification model.

2. The method for mining key tactics in wargaming simulations based on key event detection as described in claim 1, characterized in that, The process of processing the original battlefield coordination information in the wargaming data using dummy coding, interval scaling, and discretization methods to obtain coordination features includes: Dumb encoding of features is performed on the original battlefield coordination information in the wargaming data to obtain the coordination features of combat units when carrying out combat missions. The dimensionless method of interval scaling is used to process the collaborative features, transforming the differentiated format data of different combat units into unified format data; Discretize the interval-scaled collaborative features to obtain the final collaborative features.

3. The method for mining key tactics in wargaming simulations based on key event detection as described in claim 1, characterized in that, The method of increasing the dimensionality of the static features of combat units in the collaborative features and reducing the dimensionality of the dynamic features of combat units in the collaborative features includes: The static characteristics of combat units are augmented using polynomial expansion and kernel functions. Principal component analysis and linear discriminant analysis are used to reduce the dimensionality of dynamic data of combat units.

4. The method for mining key tactics in wargaming simulations based on key event detection as described in claim 1, characterized in that, The construction of a linear classification model based on the definition of novel key events includes: If the similarity between the hidden feature to be judged and the hidden feature of the known key event meets the first threshold, and the action decision corresponding to the hidden feature to be judged is different from that of the hidden feature of the known key event, and the reward obtained by the hidden feature to be judged is higher than the second threshold, then the hidden feature to be judged is a new type of key event. Alternatively, if the difference between the hidden feature to be judged and the hidden feature of the known key event meets the third threshold, and the action decision corresponding to the hidden feature to be judged and the hidden feature of the known key event is the same, and the reward obtained by the hidden feature to be judged is higher than the fourth threshold, then the hidden feature to be judged is a new type of key event.

5. The method for mining key tactics in wargaming simulations based on key event detection as described in claim 1, characterized in that, The annotation of novel key events yields a novel set of key events, including: The elements that constitute battlefield events during a limited wargaming simulation; Based on the battlefield element weights contained in the new key events output by the multi-head attention network, the core elements are determined according to the battlefield element weights, and discrete values ​​are determined for the core elements, resulting in multiple situation description keywords composed of core elements and discrete values. Based on multiple situation description keywords, a keyword template concatenation method is used to obtain the annotation information of new key events, and the annotated new key events are used as a new set of key events.

6. A device for identifying key tactics in wargaming simulations based on key event detection, comprising a processor and a memory storing a plurality of computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the key tactics mining method for wargaming process based on key event detection as described in any one of claims 1 to 5.

Images