Self-attention reinforcement learning method under incomplete information

Abstract

The application provides a self-attention reinforcement learning method under incomplete information, in the face of underwater confrontation situation complex, confrontation intelligence information is not complete, through the off-line-online hybrid driving reinforcement learning framework, through the history database introduces the confrontation expert experience of the game platform to guide the early decision of the model, and then based on the learning mode of online interaction to drive the late intelligent decision of the defense target, so as to accelerate the convergence speed of the intelligent agent and improve the intelligent decision level of the intelligent agent. Meanwhile, aiming at the multi-target attack problem, the self-attention mechanism of enemy-ally situation block interaction is proposed, the confrontation related information of the situation representation under the complete information of our side and the incomplete information of the enemy intelligence is captured, the situation information fusion representation and related information mining of the confrontation parties are realized, the situation awareness ability of the intelligent agent is improved, and the real-time and efficient decision of the underwater platform defense attacking target is realized.

Description

Technical Field

[0001] This invention belongs to the field of intelligent agents, and in particular relates to a self-attention-based reinforcement learning method under incomplete information. Background Technology

[0002] The underwater defense tactical decision-making task mainly involves the underwater platform selecting a strategy to defend against incoming targets based on the confrontation situation information, and further planning maneuver parameters based on the current strategy, repeating this process continuously until the incoming target's range is exhausted or the mission ends. Traditional decision-making methods, after target detection, select a defense strategy and set maneuver parameters based on intelligence information such as target motion elements and threat assessment. However, under conditions of target calculation non-convergence and missing intelligence information, traditional tactical decision-making methods are difficult to solve, primarily relying on commanders to infer missing information based on confrontation experience. This inevitably introduces noise and bias into the model calculation process, affecting the effectiveness of tactical defense. Therefore, in an underwater acoustic confrontation environment, how to efficiently mine global and correlated information of the adversary situation based on incomplete information to achieve real-time sequential decision-making from source data to situational awareness and underwater platform defense against incoming targets is crucial. Figure 1 The diagram illustrates the specific task modeling process for underwater defense tactical decision-making.

[0003] from Figure 1 It can be seen that the underwater defense tactical decision-making task modeled in this paper mainly includes four parts: (1) underwater situation information encoding: at the current time t, the complete information of our underwater platform is integrated and encoded. and incomplete enemy multi-target information obtained by detection ,in, D and N These refer to the feature dimension and the number of detected targets, respectively; (2) State coding: Under the influence of complex underwater environments, the underwater platform detection information includes propagation loss of the marine environment, marine reverberation, platform self-noise, radiated noise, etc., and the listening effect is related to factors such as the reflection intensity of passive sonar targets and the reflection intensity of active sonar targets. Therefore, there is a lot of missing information in the target information detected, which needs to be filled in. This paper performs zero-padding to obtain the state coding of the adversarial situation. At the same time, a state flag is introduced. In this article, boldface symbols represent matrices, and non-boldface symbols represent elements within those matrices. Record the total number of enemy and friendly forces. N+1 The first unit situation information d The complement information of the dimensional feature, where, since the information of our unit is completely known, It is D A matrix of all 1s in dimensional space, in addition to: ; in, This refers to the enemy's first i Each unit in t Time of the first d Missing dimensional features This refers to the enemy's first i Each unit in t Time of the first d Dimensional feature information; (3) Sequential decision-making of agents and experience-based decision-making of platforms: The battlefield situation is a dynamic and continuously evolving process, which is the result of dynamic game between the enemy and ourselves. The underwater confrontation situation is complex and complicated. Due to the complexity of the underwater acoustic environment, the underwater confrontation intelligence information is often incomplete, which makes the perception and actions of the enemy and ourselves full of uncertainty. This brings great difficulties to the traditional platform experience-based decision-making method that relies on information calculation and domain knowledge. Summary of the Invention

[0004] The technical problem to be solved by this invention is how to achieve more accurate and timely intelligent decision-making under the condition of incomplete intelligence information. It proposes a self-attention reinforcement learning method under incomplete information.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A self-attention-based reinforcement learning method for incomplete information includes the following steps: Step 1: Obtain situational information on both friendly and enemy underwater platforms; Step 2: Integrate and encode the situational information of both sides, including both enemy and friendly information; Step 3: Input the state-encoded situational information of both sides into the agent decision-making model; Step 4: Decode the sequential decision results of the intelligent agent and output them to the underwater platform to carry out corresponding defense actions.

[0006] Furthermore, the agent decision-making model is trained using an offline-online hybrid-driven reinforcement learning framework; The offline-online hybrid-driven reinforcement learning framework includes an agent decision-making model and an experience-based decision-making platform. The experience-based decision-making platform stores historical experience data and a historical decision database formed by real-time decision data of the agent, and trains and learns the agent decision-making model on the same time based on the data in the historical decision database.

[0007] Furthermore, the intelligent agent decision-making model includes a temporal coding module, a first self-attention mechanism module, a residual connection and normalization layer, a situational interaction module between friendly and enemy parties, a second self-attention mechanism module, and a fully connected network layer.

[0008] Furthermore, the reward function of the agent decision-making model is: ; in, These are the position coordinates of the underwater platform at time t. Represents the observed target in The detection radius at any given moment. These refer to the headings of the underwater platform and the observed target at time t, respectively. and It is the weight of the minimum safe distance between the observed target and the underwater platform and the relative hull angle, where T represents the end time of the current time window. It is the length of the time window. Represents the normalization function; This refers to the shortest distance between the underwater platform and the target's homing activation sector within a given time window during the defense and evasion process. The larger this minimum safe distance is, the greater the reward value. If the platform enters the detection sector of the observed target, the value will be negative. Empirical decision function Based on the observed target and the underwater platform The reward value for the relative hull angle at any given time is within the range of [-1, 1]. If the relative hull angle is ±90°, the value is -1, which means that the target is approaching the underwater platform in the positive direction. If this posture is maintained, the probability of detecting the underwater platform is higher and the damage from the hit is greater. If the relative hull angle is 0° or ±180°, the value is 1, which means that the observed target and the underwater platform are horizontal in this posture. If this posture is maintained, the probability of the target detecting the underwater platform will decrease and it will be difficult to hit it.

[0009] Furthermore, in step 3, when inputting the encoded situational information of both sides into the agent decision-making model, a state flag is also introduced. It records the complement information of each dimension of the situational information of a total of N+1 units, where N represents the number of enemy units being observed, and D represents the dimension of the situational information. This indicates the status flag information of our unit. These represent the state marker information of the first to Nth enemy observed target units at time t. Meanwhile, since the information of our units is completely known, Given a D-dimensional matrix of all ones, the complement information of the states of N enemy targets at time t is: ; in, This refers to the missing feature of the i-th observed enemy target at time t in the d-th dimension. This refers to the feature information of the i-th observed target unit of the enemy at time t in the d-th dimension.

[0010] Furthermore, the specific timing encoding of the timing encoding module is as follows: In our own information and enemy information Each feature representing absolute time step information is concatenated.

[0011] Furthermore, the aforementioned enemy-friendly situation segmentation and interaction module refers to the process by which the encoded enemy-friendly situation information, after passing through the temporal coding module, the first self-attention mechanism layer module, and the residual connection and normalization layer, First, based on the original situational awareness information The situation representation information is split into our complete information. Intelligence representation of incomplete information about the enemy ; Secondly, for complete characterization Connect a fully connected feedforward network to perform dimension mapping, the operation of which includes a linear transformation and a ReLU activation output: ; in, and These are the weights and biases of the fully connected feedforward network, respectively. Subsequently, the incomplete information about the enemy was characterized. as well as Input a second self-attention representation mechanism to represent the interaction between friend and foe; Finally, the interaction representation information is dimensionally mapped through a fully connected network layer.

[0012] Furthermore, the training method for the aforementioned offline-online hybrid-driven reinforcement learning framework is as follows: First, an offline policy replay pool is constructed to collect experience policy fragments under various scenarios and with intervention from experts in different domains, before training begins. This is used for expert experience guidance in the early stages of model development; Secondly, after the model begins training, sequential decision segments of the agent's decision-making model are collected simultaneously. Increasing factor In the linear incremental mode, the sequential decision segments of the agent decision model are gradually replaced by the expert experience in the experience decision platform and added to the offline policy replay pool in a first-in-first-out manner.

[0013] Furthermore, the experience-based decision-making platform refers to the platform that calculates the target based on the situational information of the incoming target, matches it with the experience database, and obtains the decision-making scheme in the experience database. If the target solution does not converge, manual intervention is introduced to perform information reasoning until a convergent defense strategy, including maneuvering, speed adjustment, and depth adjustment, and its corresponding maneuvering parameters, such as steering angle, constant speed, and constant depth, is calculated.

[0014] The present invention also provides a decision-making agent, which is trained using a self-attention-based reinforcement learning method under incomplete information. The trained decision-making agent is used in underwater platforms to make scheme decisions.

[0015] By adopting the above technical solution, the present invention has the following beneficial effects: This invention provides a self-attention-based reinforcement learning method for incomplete information situations. When facing complex underwater adversarial situations with incomplete intelligence, it employs an offline-online hybrid reinforcement learning framework. It incorporates the adversarial expert experience from a historical database to guide the model's early decisions, and then drives later intelligent decision-making based on an online interactive learning mode. This accelerates the convergence speed of the agent and improves its human-like intelligent decision-making level. Furthermore, for multi-target attack problems, it proposes a self-attention mechanism for segmented interaction of enemy and friendly situational awareness. This mechanism captures the adversarial correlation information between the situational representation under complete friendly information and the enemy intelligence under incomplete information, achieving fusion representation of situational information from both sides and mining of correlation information. This enhances the agent's situational awareness capability and enables real-time and efficient decision-making for underwater platforms defending against incoming targets. Attached Figure Description

[0016] Figure 1 A schematic diagram of a modeling framework for underwater defense tactical decision-making tasks; Figure 2 This is a schematic diagram of the intelligent agent framework for underwater defense tactical decision-making. Detailed Implementation

[0017] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0018] Figure 1 and Figure 2 This invention illustrates a specific embodiment of a self-attention-based reinforcement learning method under incomplete information, comprising the following steps: Step 1: Obtain situational information on both sides of the underwater platform.

[0019] Step 2: Integrate and encode the situational information of both sides.

[0020] In this embodiment, as Figure 1 As shown, in underwater defense tactical decision-making, after obtaining situational information from both sides, the underwater situational information needs to be integrated and encoded. At the current time t, the complete information of our underwater platform is integrated and encoded separately. and incomplete enemy multi-target information obtained by detection ,in, D and N These refer to the feature dimension and the number of detected targets, respectively. Under the influence of complex underwater environments, the detection information from underwater platforms includes propagation losses from the marine environment, ocean reverberation, platform self-noise, and radiated noise. Furthermore, the listening test results are related to factors such as the reflection intensity of passive sonar targets and the reflection intensity of active sonar targets. Therefore, the detected target information often contains missing information and needs to be completed. This embodiment uses zero-padding to obtain the state code of the adversarial situation. .

[0021] In this embodiment, in order to record the total number of enemies and allies N+1 The complement information of each dimension of the situational information of each unit is used to introduce a state label. This records which dimension of the feature was missing and then added, effectively representing the missing information in intelligence and avoiding the introduction of errors and waste of computational resources in subsequent calculations. Since the information of our unit is completely known, our state marking... It is D A 1-dimensional matrix of all ones. These represent the state marking information of the first to Nth enemy observed target units at time t, respectively. The state complement information of the N enemy observed targets at time t is as follows: ; in, This refers to the missing feature of the d-th dimension of the i-th observed target unit at time t. This refers to the feature information of the i-th observed target unit of the enemy at time t in the d-th dimension.

[0022] Step 3: Input the state-encoded situational information of both sides into the agent decision-making model; by using the agent decision-making model, a decision-making scheme can be obtained quickly.

[0023] In this embodiment, the agent decision-making model is trained using reinforcement learning within an offline-online hybrid-driven reinforcement learning framework; the agent is designed for underwater defense tactical decision-making tasks. This is achieved through a quintuple of a Markov Decision Process (MDP). To achieve decision-making task modeling, Instantiation is as follows: ① State space It is a set of enemy and friendly situational information at every point in time during a round of confrontation, in which, This refers to the state encoding at time t mentioned earlier, where T is the total number of time steps in that round; ② Action space , is the set of all possible discrete actions, where m is the total number of actions. In this embodiment, each action includes not only tactical actions for the underwater platform to defend against incoming targets, but also parameters for that action, specifically 3 depth adjustment actions, 5 speed adjustment actions, 3 steering adjustment actions, 2 active sonar switching actions, and 3 defensive actions; ③ State transition function Determined by the platform, given the current state. Actions taken The platform will then automatically calculate and provide the next state. ④ Reward function Traditional decision-making platforms provide discrete numerical reward values ​​based on clear inflection points in the situation and the results of phased confrontations. However, this method suffers from training difficulties due to the sparsity of rewards. Therefore, this paper proposes a reward function based on target distance and optimal empirical decision-making, combining domain knowledge and expert guidance. This reward function is derived from the spatiotemporal positional relationship and distance between the underwater platform and the incoming target, and is specifically defined as follows: ; in, These are the position coordinates of the underwater platform at time t. Represents the observed target in The detection radius at any given moment. These refer to the headings of the underwater platform and the observed target at time t, respectively. and It is the weight of the minimum safe distance between the observed target and the underwater platform and the relative hull angle, where T represents the end time of the current time window. It is the length of the time window. Represents the normalization function; This refers to the shortest distance between the underwater platform and the target's homing activation sector within a given time window during the defense and evasion process. The larger this minimum safe distance is, the greater the reward value. If the platform enters the detection sector of the observed target, the value will be negative. Empirical decision function Based on the observed target and the underwater platform The reward value for the relative hull angle at any given time is within the range of [-1, 1]. If the relative hull angle is ±90°, the value is -1, which means that the target is approaching the underwater platform in the positive direction. If this posture is maintained, the probability of detecting the underwater platform is higher and the damage from the hit is greater. If the relative hull angle is 0° or ±180°, the value is 1, which means that the observed target and the underwater platform are horizontal in this posture. If this posture is maintained, the probability of the target detecting the underwater platform will decrease and it will be difficult to hit it.

[0024] For the complex adversarial decision-making problem of underwater platforms defending against incoming targets, experience-based decision-making platforms rely too heavily on the commander's experience and are easily affected by missing information and computational errors. Although online decision-making by intelligent agents results in rapid strategy evolution and better adaptation to real-world environments, it faces challenges such as low data utilization and slow strategy convergence. Therefore, this embodiment mimics the way humans understand patterns by using sample mixing to achieve hybrid driving of historical databases and online interaction chains. Experience-based decision-making from the platform is introduced in the early stages of training the intelligent agent decision-making model, followed by online autonomous decision-making in the later stages. This achieves directed learning to regular fitting, thereby improving the model's convergence speed.

[0025] In this embodiment, the offline-online hybrid-driven reinforcement learning framework includes an agent decision-making model and an experience-based decision-making platform. The experience-based decision-making platform stores historical experience data and a historical decision database formed by real-time decision data of the agent, and trains and learns the agent decision-making model on the same time based on the data in the historical decision database.

[0026] In this embodiment, the training method of the offline-online hybrid-driven reinforcement learning framework is as follows: First, an offline policy replay pool is constructed to collect experience policy fragments under various scenarios and with intervention from experts in different domains, before training begins. This is used for expert experience guidance in the early stages of model development; Secondly, after the model begins training, sequential decision segments of the agent's decision-making model are collected simultaneously. Increasing factor In the linear incremental mode, the sequential decision segments of the agent decision model are gradually replaced by the expert experience in the experience decision platform and added to the offline policy replay pool in a first-in-first-out manner.

[0027] The sequential decision segments of the agent decision-making model are gradually added to the offline policy replay pool, replacing the expert experience in the experience-based decision-making platform in a first-in-first-out manner. This achieves a hybrid driving method that combines historical databases and online interaction chains from a sample mixing perspective: , This refers to each A fixed time step at a rate Add sequential decision fragments to the agent decision-making model.

[0028] This embodiment uses a linear incremental method to guide the mixing of two types of samples in different proportions at different training stages, thereby leading to a change in the sampling method. Different sample sampling methods are adopted in different policy learning stages to improve algorithm performance and data utilization from the perspective of sample sampling method, and further improve the performance of offline-online hybrid reinforcement learning algorithm.

[0029] The aforementioned experience-based decision-making platform refers to the platform that calculates the target based on the situational information of the incoming target, then matches it with an experience database to obtain a decision-making scheme from the experience database. If the target solution does not converge, manual intervention is introduced to perform information reasoning until a convergent defense strategy, including maneuvering, speed adjustment, and depth adjustment, and its corresponding maneuvering parameters, such as steering angle, constant speed, and constant depth, is calculated.

[0030] In this embodiment, as Figure 2 The agent decision-making model shown includes a temporal coding module, a first self-attention mechanism module, a residual connection and normalization layer, a situational interaction module between friendly and enemy parties, a second self-attention mechanism module, and a fully connected network layer.

[0031] In this embodiment, considering that the self-attention mechanism does not include temporal processing methods such as convolution and recursion, a temporal encoding module is introduced to temporally encode the situational information in order to fully utilize the temporal characteristics of the situational information during adversarial processes. This module injects absolute time step information about the sequence markers into the obtained situational information.

[0032] The specific timing encoding of the timing encoding module is as follows: in one's own information and enemy information Each feature representing an absolute time step is concatenated, increasing its dimension by 1. The purpose of introducing an absolute time step is to enable the agent's decision-making model to know the chronological sequence of data, thus fully utilizing the temporal information of the platform during adversarial processes.

[0033] Furthermore, considering the incomplete information regarding multiple enemy observation targets, a first self-attention mechanism module is introduced to effectively mine the enemy's current situational information, thereby effectively mining the correlation between multi-source information inputs. Specifically, the semantic features mined from the adversarial situational information obtained at the current moment are: ; in, This refers to Query( ) and Key( ) dimensions All of these are temporal input information after sequential encoding. They are the three specified inputs Q (query), K (key), and V (value) required for attention calculation. Then, the attention calculation result is obtained through a formula.

[0034] Clearly, introducing self-attention makes it easier to capture long-term temporally dependent features in situational information. This is because RNNs or LSTMs require sequential computation, accumulating information over several time steps to link distant dependent features, and the greater the distance, the lower the likelihood of effective capture. However, self-attention directly links any two pieces of information in the situational information in a single computational step, significantly reducing the distance between distant dependent features and facilitating their effective utilization.

[0035] Then, after the information from the first self-attention mechanism module is processed by residual connection and layer normalization, a block interaction module for the situation of both sides is introduced to address the information representation correlation between the adversaries.

[0036] In this embodiment, the enemy and friendly situation segmentation and interaction module refers to the module that, after the encoded enemy and friendly situation information passes through the temporal coding module, the first self-attention mechanism layer module, and the residual connection and normalization layer, ... First, based on the original situational awareness information The situation representation information is split into our complete information. Intelligence representation of incomplete information about the enemy ; Secondly, for complete characterization Connect a fully connected feedforward network to perform dimension mapping, the operation of which includes a linear transformation and a ReLU activation output: ; in, and These are the weights and biases of the fully connected feedforward network, respectively. Subsequently, the incomplete information about the enemy was characterized. as well as Input a second self-attention representation mechanism to represent the interaction between friend and foe; ; in, and .

[0037] Finally, the interaction representation information is dimensionally mapped through a fully connected network layer.

[0038] Step 4: Decode the sequential decision-making results of the agent and output them to the underwater platform for corresponding defensive actions. Decode the agent's decisions and output the current strategy in real time. and the corresponding parameters It is used to assist commanders in making command decisions.

[0039] The present invention also provides a decision-making agent, which is trained using a self-attention-based reinforcement learning method under incomplete information. The trained decision-making agent is used in underwater platforms to make scheme decisions.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A self-attention-based reinforcement learning method under incomplete information, characterized in that, Includes the following steps: Step 1: Obtain situational information on both friendly and enemy underwater platforms; Step 2: Integrate and encode the situational information of both sides, including both enemy and friendly information; Step 3: Input the state-encoded situational information of both sides into the agent decision-making model; The agent decision-making model includes a temporal coding module, a first self-attention mechanism module, a residual connection and normalization layer, a situational interaction module between friendly and enemy sides, a second self-attention mechanism module, and a fully connected network layer. The reward function of the agent decision-making model is: ; in, These are the position coordinates of the underwater platform at time t. Represents the observed target in The detection radius at any given moment. These refer to the headings of the underwater platform and the observed target at time t, respectively. and It is the weight of the minimum safe distance between the observed target and the underwater platform and the relative hull angle, where T represents the end time of the current time window. It is the length of the time window. Represents the normalization function; This refers to the shortest distance between the underwater platform and the target's homing activation sector within a given time window during the defense and evasion process. The larger this minimum safe distance is, the greater the reward value. If the platform enters the detection sector of the observed target, the value will be negative. Empirical decision function Based on the observed target and the underwater platform The reward value for the relative angle at any given time is within the range of [-1, 1]. If the relative angle is ±90°, the value is -1, which means that the target is approaching the underwater platform in the positive direction. If this posture is maintained, the probability of detecting the underwater platform is higher and the damage from the hit is greater. If the relative angle is 0° or ±180°, the value is 1, which means that the observed target and the underwater platform are horizontal in this posture. If this posture is maintained, the probability of the target detecting the underwater platform will decrease and it will be difficult to hit it. When inputting the state-encoded situational information of both friendly and enemy sides into the agent decision-making model, a state label is also introduced. It records the complement information of each dimension of the situational information of a total of N+1 units, where N represents the number of enemy units being observed, and D represents the dimension of the situational information. This indicates the status flag information of our unit. These represent the state marker information of the first to Nth enemy observed target units at time t. Meanwhile, since the information of our units is completely known, Given a D-dimensional matrix of all ones, the complement information of the states of N enemy targets at time t is: ； in, This refers to the missing feature of the i-th observed enemy target at time t in the d-th dimension. This refers to the feature information of the i-th observed target unit of the enemy at time t in the d-th dimension; The specific timing encoding of the timing encoding module is as follows: in one's own information and enemy information Each feature representing absolute time step information is concatenated; The aforementioned enemy and friendly situation segmentation and interaction module refers to the process by which the encoded enemy and friendly situation information, after passing through the temporal coding module, the first self-attention mechanism layer module, and the residual connection and normalization layer, ... First, based on the original situational awareness information The situation representation information is split into our complete information. Intelligence representation of incomplete information about the enemy ; Secondly, for complete characterization Connect a fully connected feedforward network to perform dimension mapping, the operation of which includes a linear transformation and a ReLU activation output: ； in, and These are the weights and biases of the fully connected feedforward network, respectively. Subsequently, the incomplete information about the enemy was characterized. as well as Input a second self-attention representation mechanism to represent the interaction between friend and foe; Finally, the interaction representation information is dimension-mapped through a fully connected network layer; Step 4: Decode the sequential decision results of the intelligent agent and output them to the underwater platform to carry out corresponding defense actions.

2. The method according to claim 1, characterized in that, The agent decision-making model is trained using an offline-online hybrid-driven reinforcement learning framework. The offline-online hybrid-driven reinforcement learning framework includes an agent decision-making model and an experience-based decision-making platform. The experience-based decision-making platform stores historical experience data and a historical decision database formed by real-time decision data of the agent, and trains and learns the agent decision-making model on the same time based on the data in the historical decision database.

3. The method according to claim 2, characterized in that, The training method of the offline-online hybrid-driven reinforcement learning framework is as follows: First, an offline policy replay pool is constructed to collect experience policy fragments under various scenarios and with intervention from experts in different domains, before training begins. This is used for expert experience guidance in the early stages of model development; Secondly, after the model begins training, sequential decision segments of the agent's decision-making model are collected simultaneously. Increasing factor In the linear incremental mode, the sequential decision segments of the agent decision model are gradually replaced by the expert experience in the experience decision platform and added to the offline policy replay pool in a first-in-first-out manner.

4. The method according to claim 3, characterized in that, The aforementioned experience-based decision-making platform refers to the platform that calculates the target based on the situational information of the incoming target, then matches it with an experience database to obtain a decision-making scheme from the experience database. If the target solution does not converge, manual intervention is introduced to perform information reasoning until a convergent defense strategy, including maneuvering, speed adjustment, and depth adjustment, and its corresponding maneuvering parameters are calculated. The maneuvering parameters include steering angle, constant speed, and constant depth.

Images