Bayesian neural network-based auxiliary decision recommendation model adaptive method and system

Abstract

The invention relates to the technical field of air combat intelligent aid decision making, in particular to an aid decision making recommendation model adaptive method and system based on a Bayesian neural network, and is particularly suitable for man-machine collaborative decision making in a 2V2 air combat game scene. By designing a self-adaptive framework and an incremental learning algorithm, the model can be dynamically updated based on real-time data in the use process of a user, so that the recommendation accuracy and robustness are improved. The method comprises the following steps: constructing an incremental learning algorithm of a Bayesian neural network, analyzing the time complexity and the space complexity of the incremental learning algorithm, evaluating the real-time performance of the incremental learning algorithm in an online reasoning process, and researching the robustness performance under the condition of data uncertainty. The method can be widely applied to strategy selection in an intelligent recommendation system, an auxiliary decision-making platform and a multi-agent game.

Description

Technical Field

[0001] This invention belongs to the field of intelligent auxiliary decision-making technology in air combat, and particularly relates to an adaptive method and system for auxiliary decision recommendation model based on Bayesian neural network, which is especially suitable for human-machine collaborative decision-making in 2V2 air combat game scenarios. Background Technology

[0002] Existing recommendation decision-making models are mostly based on deep neural networks or collaborative filtering methods, but they lack adaptability when flight data changes dynamically. Especially when facing uncertain data, traditional methods struggle to guarantee robustness. Bayesian Neural Networks (BNNs), by introducing probability distributions into the weights, can naturally handle uncertainty. However, existing BNN recommendation models are mostly statically trained and lack online adaptive update mechanisms, limiting their application in real-time decision-making. Summary of the Invention

[0003] Purpose of the invention: To propose an adaptive strategy for an auxiliary recommendation decision model based on Bayesian neural networks. By introducing incremental learning and an adaptive framework, the model can be continuously updated during user interaction, thereby improving the real-time performance, robustness, and computational efficiency of the system.

[0004] Technical solution: An adaptive method for an auxiliary decision recommendation model based on Bayesian neural networks includes the following steps: (1) Obtain online interaction data from pilots to form an input dataset. ,in For feature vectors, For corresponding recommendation tags or decision feedback; (2) Construct a Bayesian neural network model, its parameters Follows prior distribution Through posterior distribution Update the parameters; (3) New data and historical data Perform concept drift detection based on the KS test; for each key feature dimension in the flight data Perform the KS test separately: Let Representing historical data In the Let the empirical distribution function on each feature be... Indicates new data In the Calculate the KS statistic based on the empirical distribution function of each feature:

[0005] If the average KS statistic of all features Exceeding the set threshold Then, an incremental learning strategy is adopted, utilizing newly arrived data. Execute the adaptive incremental learning module; (4) During user usage, incremental updates are performed using variational inference methods to calculate the true posterior distribution. The approximation problem is transformed into finding the optimal variational parameters. The optimization problem, where the variational distribution Modeling is performed using a family of Gaussian distributions; optimal variational parameters for:

[0006] in, For dynamic regularization coefficients, It is the log-likelihood of the new data. Represents the expectation under the variational distribution. It is the Kullback-Leibler divergence between the new variational distribution and the old variational distribution.

[0007] Furthermore, the method also includes: (5) in the recommendation process, performing computational complexity, real-time performance and robustness analysis on the model to ensure the reliability and interpretability of the recommendation.

[0008] Furthermore, the computational complexity analysis includes the derivation of time and space complexity, wherein the time complexity is: in Indicates the number of samples. Indicates the number of layers in a neural network. This represents the average dimension of each layer; the space complexity is: .

[0009] Furthermore, the incremental learning algorithm is based on variational inference, and its optimization objective function is: .

[0010] Furthermore, the variational parameter update strategy aims to find the optimal variational parameters. The optimization process is implemented using gradient descent-like algorithms; learning rate This acts as the step size, and the specific update formula is as follows:

[0011] in, It is the first Variational parameters at the next iteration Is the objective function in gradient at, It is a dynamic learning rate calculated based on the degree of concept drift. These are the updated variational parameters.

[0012] Furthermore, the learning rate Dynamically adjust based on the degree of concept drift:

[0013] in, The base learning rate is set to 0.001 by default. To adjust the sensitivity coefficient, set it to 1.2; The concept drift detection threshold is set to 0.15.

[0014] Furthermore, the posterior distribution is as follows: .

[0015] An adaptive system for decision-aid recommendation models based on Bayesian neural networks, the system comprising: The data acquisition module is used to collect user online interaction data; The model training module is used to perform Bayesian neural network training and incremental updates. The performance analysis module is used for computational complexity, real-time performance, and robustness analysis. The recommendation output module is used to generate the final recommendation results; the encapsulation and interface module is used to implement online calling and service of the model.

[0016] Beneficial effects: (1) Strong adaptive update capability: The Bayesian neural network-based auxiliary recommendation decision model proposed in this invention can realize real-time adaptive update of model parameters using online data, breaking through the limitations of traditional recommendation systems that rely on offline batch training and are difficult to adapt to dynamic environmental changes. By introducing an incremental learning mechanism, this invention can quickly absorb new information while maintaining historical knowledge, avoiding the forgetting effect.

[0017] (2) Uncertainty Quantification and Robustness Enhancement: Compared to traditional neural networks that only output point estimates, this invention uses Bayesian inference to probabilistically model network weights, which can explicitly quantify the uncertainty of prediction results. When faced with data noise, missing data, and distribution drift, this invention maintains higher robustness and improves the reliability of recommendation decisions.

[0018] (3) Enhanced interpretability: This invention utilizes posterior distributions. The estimation ensures that the recommendation results not only contain predicted values ​​but also confidence intervals or variance information, providing interpretable evidence for users and decision-makers. This feature has significant application value in high-risk scenarios such as financial risk control and medical diagnosis.

[0019] (4) Controllable computational complexity: In designing the incremental learning algorithm, this invention employs methods such as variational inference (VI) and stochastic gradient Markov chain Monte Carlo (SG-MCMC) to keep the computational complexity of updates at the polynomial level. For example, the time complexity of this invention is... ,in For sample batch size, The number of network layers is O(n); the space complexity is O(n).

[0020] (5) Real-time performance meets online requirements: among which This refers to the number of neurons per layer. Compared to traditional full retraining methods, this invention significantly reduces computational costs, keeping the single prediction time within [a certain range]. It has the capability to respond to user requests in milliseconds, meeting the application requirements of real-time recommendation scenarios.

[0021] (6) The system is easy to encapsulate and call: This invention not only proposes an algorithm framework, but also provides an interface encapsulation mechanism to support calls in a multi-platform environment. Users can directly access the adaptive recommendation service through the API interface, reducing deployment and usage costs. Attached Figure Description

[0022] Figure 1 This is the overall flowchart of the present invention.

[0023] Figure 2 This is the KS test chart of the present invention.

[0024] Figure 3 This is a graph showing the change of the objective function of this invention.

[0025] Figure 4 This is a comparison chart of the model performance before and after the adaptive update of the present invention.

[0026] Figure 5 This is a robustness analysis diagram of the present invention. Detailed Implementation

[0027] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0028] In this application, the terms "upper," "lower," "left," "right," "front," "rear," "top," "bottom," "inner," "outer," "middle," "vertical," "horizontal," "lateral," and "longitudinal" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are primarily for the purpose of better describing this application and its embodiments, and are not intended to limit the indicated device, element, or component to having a specific orientation, or to be constructed and operated in a specific orientation.

[0029] Furthermore, the terms "installation," "setup," "equipped with," "connection," "linking," and "socketing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral structure; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium, or an internal connection between two devices, components, or parts. Those skilled in the art can understand the specific meaning of these terms in this application based on the specific circumstances.

[0030] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0031] An adaptive method for an auxiliary decision-making recommendation model based on Bayesian neural networks includes the following steps: (1) Obtain online interaction data from pilots to form an input dataset. ,in This is a feature vector for flight data, including relative situational features such as relative speed, altitude, and distance. For the corresponding recommendation tags or decision feedback, For data scale; (2) Construct a Bayesian neural network model, its parameters Follows prior distribution Through posterior distribution

[0032] Update the parameters, where the parameters are... These represent the weights and biases for each level of the network; (3) Utilize real-time flight data for concept drift detection. New flight data will be continuously generated during user operation. First, regarding the new data and historical data Concept drift detection based on the Kolmogorov-Smirnov (KS) test was performed to scientifically determine whether significant changes had occurred in the data distribution. This was done for each key feature dimension in the flight data. (e.g., relative speed, height, distance, etc.), perform KS test respectively: Let Representing historical data In the Let the empirical distribution function on each feature be... Indicates new data In the Calculate the KS statistic based on the empirical distribution function of each feature:

[0033] in Indicates all possible values Find the supremum (i.e., the maximum vertical distance between two empirical distribution functions).

[0034] (4) Incremental updates are performed using variational inference methods to obtain the true posterior distribution. The approximation problem is transformed into finding the optimal variational parameters. The optimization problem, where the variational distribution A Gaussian distribution family is used for modeling. Optimal variational parameters. for:

[0035] in, This is a dynamic regularization coefficient used to balance the learning of new knowledge with the retention of old knowledge, based on new data. Compared with historical data The distribution similarity is adaptively adjusted: when the statistic When it is greater than 0.25, set when When the value is between 0.1 and 0.25, set when When less than 0.1, set . It is the log-likelihood of the new data. Represents the expectation under the variational distribution. It is the Kullback-Leibler divergence between the new variational distribution and the old variational distribution. Adding it to the optimization objective function helps prevent the model from deviating too much from historical knowledge and controls the forgetting rate.

[0036] (5) Adaptive adjustment of learning rate, learning rate Dynamically adjust based on the degree of concept drift:

[0037] in, The base learning rate is set to 0.001 by default. To adjust the sensitivity coefficient, set it to 1.2; The concept drift detection threshold is set to 0.15.

[0038] (6) Based on the updated model, for new user requests Perform prediction and recommendation, and output the expected value:

[0039] (7) During the recommendation process, the computational complexity, real-time performance and robustness of the model are analyzed to ensure the reliability and interpretability of the recommendations. This invention proposes an adaptive method for an auxiliary decision-making recommendation model based on Bayesian Neural Network (BNN). The core idea is to introduce uncertainty modeling into the parameter space of the traditional neural network, and to characterize the uncertainty of the model in response to data, environment and user behavior through the probability distribution of the parameters, thereby improving the robustness and adaptability of the recommendation decision.

[0040] Combination Figure 1 The overall system architecture and main steps include: 1. Data Acquisition and Preprocessing Module: Used to collect user interaction data, external environment data, and task constraints in real time, and to perform normalization, feature extraction, and dimensionality compression. The main function of this module is to acquire and process data from multiple sources in real time, ensuring that the subsequent Bayesian neural network-assisted decision-making recommendation model can operate with high-quality data support. Specifically, the module first collects users' historical behavior data, real-time operation records, and preference feedback through a user interaction interface. Then, the module performs preprocessing operations on the raw data, including data cleaning, missing value imputation, and outlier detection, to improve the accuracy and consistency of the data. After initial cleaning, the system further normalizes the data to eliminate differences between different dimensions and units, ensuring that the model input has a uniform scale.

[0041] The acquired online interaction data from pilots will form the input dataset. ,in The flight data feature vector includes four relative situational features: relative speed, relative altitude, relative distance, and relative heading angle. For the corresponding recommendation tags or decision feedback, For data scale; 2. Bayesian Neural Network Modeling Module: This module is responsible for building the core neural network structure of the decision-making recommendation model and modeling the weight parameters as probability distributions to quantify uncertainty. First, the module constructs a neural network structure with input, intermediate hidden, and output layers based on preprocessed feature data, and sets an initial prior distribution (such as a Gaussian distribution) for the weight parameters of each layer. Then, the network is optimized using the training dataset, and the posterior distribution of the weights is approximated using variational inference (VI) or Markov chain Monte Carlo (MCMC) methods to obtain information about the uncertainty of the network parameters. During model training, the module optimizes the loss function, including maximizing the likelihood function and minimizing the KL divergence term, to ensure the model maintains robust predictive ability under different data distributions. This module also supports outputting the predicted distribution, not just point estimates, providing a confidence reference for subsequent decisions.

[0042] Step 2.1 Let the input features be... The output recommendation decision is The goal of BNN is to learn conditional probability distributions:

[0043] in, These are the weights and bias parameters of the neural network. This is the training dataset.

[0044] Step 2.2 Untested Distribution and Posterior Inference: (For parameters) Assign a Gaussian prior distribution:

[0045] In observation data Below, the posterior distribution is

[0046] in:

[0047] Since the integral is not analytical, this invention uses variational inference (VI) for approximation.

[0048] Step 2.3 Variational Inference: Introducing Variational Distribution Approximate posterior The optimization objective is to minimize the Kullback-Leibler (KL) divergence.

[0049] This is equivalent to maximizing the Evidence Lower Bound (ELBO).

[0050] Step 2.4 Online Inference and Uncertainty Quantification: During prediction, the recommended result is as follows:

[0051] Prediction uncertainty is given by the variance of the parameter distribution:

[0052] This uncertainty index can be used to estimate decision confidence and improve system robustness.

[0053] 3. Concept Drift Detection Module Based on KS Test: This module is responsible for scientifically detecting changes in data distribution, providing a basis for adaptive updates. The module calculates the KS statistic for the distribution of old and new data for each key flight feature, and uses this to comprehensively determine the occurrence and extent of concept drift.

[0054] New flight data will be continuously generated during user operation. First, regarding the new data and historical data Concept drift detection based on the Kolmogorov-Smirnov (KS) test was performed to scientifically determine whether significant changes had occurred in the data distribution. This was done for each key feature dimension in the flight data. (e.g., relative speed, height, distance, etc.), perform KS test respectively: Let Representing historical data In the Let the empirical distribution function on each feature be... Indicates new data In the Calculate the KS statistic based on the empirical distribution function of each feature:

[0055] in Indicates all possible values The supremum (i.e., the maximum vertical distance between two empirical distribution functions) is calculated. The KS statistic for the four characteristic dimensions is as follows: Figure 2 As shown, if the average KS statistic of all features Exceeding the set threshold Then, an incremental learning strategy is adopted, utilizing newly arrived data. Execute the adaptive incremental learning module.

[0056] 4. Adaptive Incremental Learning Module: The core function of this module is to continuously and stably update the Bayesian neural network online during realistic air combat training and missions. This allows the model to learn from the pilot's actual decision-making behavior and gradually evolve into an intelligent assistant highly compatible with the specific pilot's decision-making style. The module receives real-time data streams from the air combat simulation platform, including aircraft status, battlefield situation, and the pilot's final decision actions. Through incremental learning algorithms, new decision-making patterns are integrated into the model while preserving existing air combat knowledge.

[0057] Step 4.1 Specific Implementation of the Incremental Update Mechanism The incremental update mechanism is implemented in engineering using variational inference methods. Specifically, it involves calculating the true posterior distribution. The approximation problem is transformed into finding the optimal variational parameters. The optimization problem, where the variational distribution The Gaussian distribution family is used for modeling.

[0058] The objective function to be optimized is:

[0059] Optimal variational parameters for:

[0060] in, This is a dynamic regularization coefficient used to balance the learning of new knowledge with the retention of old knowledge, based on new data. Compared with historical data The distribution similarity is adaptively adjusted: when the statistic When it is greater than 0.25, set when When the value is between 0.1 and 0.25, set when When less than 0.1, set . It is the log-likelihood of the new data. Represents the expectation under the variational distribution. This is the Kullback-Leibler divergence between the new and old variational distributions. Adding it to the objective function helps prevent the model from deviating excessively from historical knowledge and controls the forgetting rate. The curve showing the change of the objective function during incremental learning is as follows: Figure 3 As shown.

[0061] Step 4.2 Detailed Implementation of the Adaptive Strategy Framework 4.2.1 Dynamic Learning Rate Adjustment Strategy Learning rate Dynamically adjust based on the degree of concept drift:

[0062] in, The base learning rate is set to 0.001 by default. To adjust the sensitivity coefficient, set it to 1.2; The concept drift detection threshold is set to 0.15.

[0063] 4.2.2 Variational Parameter Update Strategy In step 4.1, we defined the objective function to be minimized and sought the optimal variational parameters. This optimization process is implemented using gradient descent-like algorithms. The learning rate obtained above... This acts as the step size, and the specific update formula is as follows:

[0064] in, It is the first Variational parameters in the next iteration. Is the objective function in The gradient at that point. It is the dynamic learning rate that we calculate based on the degree of concept drift. These are the updated variational parameters. The changes in model accuracy and Kappa coefficient before and after the adaptive update are as follows: Figure 4 As shown.

[0065] 5. Performance Analysis Module: This module is responsible for comprehensively evaluating the operational efficiency and reliability of the adaptive Bayesian neural network model. Specifically, it includes three aspects: computational complexity analysis, real-time performance analysis, and robustness analysis. Computational complexity analysis assesses the model's processing capabilities at different network sizes by statistically analyzing the computational costs of forward propagation, backpropagation, and parameter updates. Real-time performance analysis focuses on the latency of the model during online inference and incremental updates, ensuring that the system can complete recommendations within milliseconds or an acceptable latency range. Robustness analysis utilizes experiments with prediction uncertainty, data noise, and outlier samples to evaluate the model's stability and reliability when facing distribution drift or anomalous data. The analysis results of this module can be used to guide parameter optimization and system resource scheduling in the incremental learning module.

[0066] Step 5.1 Output Time Complexity: The time complexity of each update in variational inference is:

[0067] in For batch size, For input dimensions, Let be the size of the hidden layer. Incremental learning strategies can reduce the complexity to: [formula missing - likely a number] by limiting the range of updated parameters.

[0068] Step 5.2 Output Space Complexity: BNN needs to store the mean and variance of the parameter distribution, and the space complexity is: in This represents the total number of parameters.

[0069] Step 5.3 Output Online Inference Time and Model Update Time: This step aims to accurately quantify and monitor the real-time performance of the decision support and recommendation system, ensuring that the system meets the stringent timeliness requirements of an air combat environment. After each adaptive update, the system automatically calculates and outputs two key performance indicators: online inference time and model update time. For low-latency scenarios, experiments show that online inference and model adaptive updates can be completed within 0.5 seconds, meeting the needs of real-time recommendation.

[0070] Step 5.4 Robustness Analysis: In the original feature vector Superimposed Gaussian noise Observe the changes in the model output. Calculate the prediction bias.

[0071] And the decision reversal rate (the proportion of samples where the recommended decision is fundamentally changed (e.g., from "attack" to "avoidance"). The output variance and decision reversal rate curves for different noise levels are shown below. Figure 5 As shown.

[0072] 5. Online Invocation and Encapsulation Interface: This module transforms the prediction results processed by the Bayesian neural network into actual recommendation strategies and decision-making schemes. Based on the model's output prediction distribution and uncertainty index, the module calculates the expected return and risk level of each candidate operation, and generates optimal or near-optimal decision recommendations according to task constraints. Output results may include specific recommendation actions, ranking lists, confidence scores, and auxiliary prompts, and are pushed to end users or control systems in real time through user interfaces or system interfaces. Furthermore, the module supports a result feedback mechanism, re-feeding user adoption status or actual execution effects back to the data acquisition module, forming a closed-loop adaptive learning system.

[0073] The above description is merely a specific embodiment of the present invention, providing a detailed description of the invention. Parts not covered herein are conventional techniques. However, the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An adaptive method for an auxiliary decision-making recommendation model based on Bayesian neural networks, characterized in that, Includes the following steps: (1) Obtain online interaction data from pilots to form an input dataset. ,in For feature vectors, For corresponding recommendation tags or decision feedback; (2) Construct a Bayesian neural network model, its parameters Follows prior distribution Through posterior distribution Update the parameters; (3) New data and historical data Perform concept drift detection based on the KS test; for each key feature dimension in the flight data Perform the KS test separately: Let Representing historical data In the The empirical distribution function on each feature, let Indicates new data In the Calculate the KS statistic based on the empirical distribution function of each feature: If the average KS statistic of all features Exceeding the set threshold Then, an incremental learning strategy is adopted, utilizing newly arrived data. Execute the adaptive incremental learning module; (4) During user usage, incremental updates are performed using variational inference methods to calculate the true posterior distribution. The approximation problem is transformed into finding the optimal variational parameters. The optimization problem, where the variational distribution Modeling is performed using a family of Gaussian distributions; optimal variational parameters for: in, For dynamic regularization coefficients, It is the log-likelihood of the new data. Represents the expectation under the variational distribution. It is the Kullback-Leibler divergence between the new variational distribution and the old variational distribution.

2. The method according to claim 1, characterized in that, The method further includes: (5) In the recommendation process, performing computational complexity, real-time performance and robustness analysis on the model to ensure the reliability and interpretability of the recommendation.

3. The method according to claim 2, characterized in that, The computational complexity analysis includes the derivation of time and space complexity, where the time complexity is: in Indicates the number of samples. Indicates the number of layers in a neural network. This represents the average dimension of each layer; the space complexity is: .

4. The method according to claim 3, characterized in that, The incremental learning algorithm is based on variational inference, and its objective function is: 。 5. The method according to claim 4, characterized in that, The variational parameter update strategy aims to find the optimal variational parameters. The optimization process is implemented using gradient descent-like algorithms; learning rate This acts as the step size, and the specific update formula is as follows: in, It is the first Variational parameters at the next iteration Is the objective function in gradient at, It is a dynamic learning rate calculated based on the degree of concept drift. These are the updated variational parameters.

6. The method according to claim 5, characterized in that, The learning rate Dynamically adjust based on the degree of concept drift: in, The base learning rate is set to 0.001 by default. To adjust the sensitivity coefficient, set it to 1.2; The concept drift detection threshold is set to 0.

15.

7. The method according to claim 1, characterized in that, The posterior distribution is as follows: 。 8. An adaptive system for an auxiliary decision-making recommendation model based on a Bayesian neural network, used to implement the method according to any one of claims 1-7, characterized in that: The system includes: The data acquisition module is used to collect user online interaction data; The model training module is used to perform Bayesian neural network training and incremental updates. The performance analysis module is used for computational complexity, real-time performance, and robustness analysis. The recommendation output module is used to generate the final recommendation results; The encapsulation and interface modules are used to enable online invocation and service of the model.

Images