Confidence-based multi-agent system interpretability enhancement method, system and equipment and medium

Abstract

The invention discloses a confidence-based multi-agent system interpretability enhancement method, system and device and a medium, and belongs to the technical field of artificial intelligence, and the method comprises the steps: collecting various information of each agent in a multi-agent system, and carrying out the preprocessing to obtain the preprocessed information; receiving the preprocessed information based on a deep learning model, and outputting a confidence value; correcting a confidence coefficient value in combination with environment change and information interaction between intelligent agents; a visual interface for displaying confidence is constructed, and the confidence calculation basis and time change of specific agent decision making are supported to be checked; the strategy of a multi-agent algorithm is adjusted according to the confidence coefficient of agent decision, the confidence coefficient is introduced in the training process to serve as a training target, and a cooperation mechanism between agents is optimized by utilizing confidence coefficient information. According to the method, the interpretability of the decision-making process of the multi-agent system is effectively improved, the algorithm performance is optimized, and the collaboration capability among the agents is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a confidence-based multi-agent system explainability enhancement method, system, device and medium. BACKGROUND

[0002] In today's era of rapid technological advancement and artificial intelligence development like a raging fire, Multi-Agent System (MAS) is like a shining star, with its unique advantages and powerful functions, shining brightly in many fields. Multi-Agent Reinforcement Learning (MARL) is a key technology of multi-agent system, its principle is to learn through the interaction between multiple agents and the environment, so as to gradually achieve the completion of complex tasks. In this process, the agent constantly tries various actions and adjusts its behavior strategy according to the feedback from the environment. However, as the application scenarios of multi-agent system become increasingly complex and diverse, the opacity and unexplainability of multi-agent reinforcement learning algorithms like hidden stumbling blocks are increasingly exposed.

[0003] Because the decision-making mechanism inside the multi-agent reinforcement learning algorithm often involves complex mathematical models and a large amount of data calculation, it is difficult for ordinary users and even some professional researchers to understand how the algorithm makes decisions. This opacity seriously limits the user's understanding and trust of the algorithm's decision-making process. In some critical fields with high requirements for safety and reliability, such as medical surgery assistance, aerospace control, etc., users are difficult to entrust important tasks to these algorithms for execution if they cannot understand the basis of the algorithm's decision, thereby hindering the further promotion and application of multi-agent reinforcement learning algorithms in these fields.

[0004] Currently, in the evaluation of multi-agent reinforcement learning algorithms, most research focuses on the performance and scalability of the algorithm. Researchers strive to improve the calculation speed of the algorithm, optimize the decision results, and enable the algorithm to adapt to more agents and more complex environments. However, they often overlook the critical issues of training cost and data efficiency in actual distributed deployment. In a complex and changing real environment, training an efficient multi-agent reinforcement learning algorithm requires a large amount of computing resources and time, and also needs to handle massive data. If the training cost and data efficiency are not considered, even if the algorithm has good performance and scalability in theory, it may not be able to effectively play its role in practical applications due to excessive resource consumption and long training time.

[0005] In the field of algorithm interpretability research, although many scholars have attempted to find solutions, current research mostly focuses on single-agent systems. Single-agent systems are relatively simple, and their decision-making processes are relatively easy to understand and explain. However, multi-agent systems involve complex interactions and collaborative mechanisms between multiple agents, making the interpretability problem of multi-agent reinforcement learning algorithms more complex and challenging. Existing interpretability methods often fall short when faced with complex game environments and intricate interactions between multiple agents, failing to provide comprehensive and effective explanations.

[0006] Therefore, introducing an effective interpretability method into multi-agent systems, particularly leveraging the concept of confidence to enhance understanding of the algorithm's decision-making process, has become a crucial problem urgently needing to be solved in the field of artificial intelligence. Confidence provides a quantitative reliability indicator for the algorithm's decisions, allowing users to clearly understand the credibility of the decisions made by the algorithm. By introducing confidence, we can gain a deeper understanding of the algorithm's decision-making basis, increase user trust in the algorithm, and thus promote the widespread application of multi-agent reinforcement learning algorithms in more key areas. Summary of the Invention

[0007] In view of the above-mentioned problems, the present invention is proposed.

[0008] Therefore, the technical problem solved by this invention is: how to address the issue that existing interpretability methods often fall short in providing comprehensive and effective explanations when faced with complex game environments and intricate interactions between multiple agents.

[0009] To address the aforementioned technical problems, this invention provides the following technical solution: a method for enhancing the interpretability of a multi-agent system based on confidence, comprising: collecting various information from each agent in the multi-agent system and preprocessing it to obtain preprocessed information; receiving the preprocessed information based on a deep learning model and outputting confidence values; correcting the confidence values ​​by combining environmental changes and information interaction between agents; constructing a visual interface to display the confidence, supporting the viewing of the basis for confidence calculation and time changes of specific agent decisions; adjusting the strategy of the multi-agent algorithm according to the confidence of agent decisions, introducing confidence as a training objective during training, and utilizing confidence information to optimize the cooperation mechanism between agents.

[0010] As a preferred embodiment of the confidence-based multi-agent system interpretability enhancement method described in this invention, the step of collecting various information from each agent in the multi-agent system and performing preprocessing to obtain preprocessed information includes: collecting various information from multiple agents during the execution process; performing integrity checks and format standardization on the collected information; and preprocessing the formatted information to obtain preprocessed information.

[0011] As a preferred embodiment of the confidence-based multi-agent system interpretability enhancement method described in this invention, the step of receiving preprocessed information and outputting confidence values ​​based on a deep learning model includes: feeding the preprocessed information into a deep learning model substructure for processing image information and sequence information respectively; extracting image features and temporal features and fusing them into a set of joint feature vectors; and inputting the joint feature vectors into a fully connected network structure to generate confidence output values.

[0012] As a preferred embodiment of the confidence-based multi-agent system interpretability enhancement method described in this invention, the step of correcting the confidence value by combining environmental changes and information interaction between agents includes: acquiring current environmental change information and information interaction data between agents; determining whether environmental changes may affect the decision confidence; calling a preset confidence correction function based on environmental change information and interaction data to calculate the confidence correction value; and applying the correction value to the confidence output value to generate the corrected confidence result.

[0013] As a preferred embodiment of the confidence-based multi-agent system interpretability enhancement method described in this invention, the step of feeding the preprocessed information into a deep learning model substructure for processing image information and sequence information includes: a convolutional neural network for processing image-type observation data, performing multiple convolution and pooling operations sequentially on the input image to generate an image feature map; a recurrent neural network for processing time-series observation data, extracting hidden state information at each time step in chronological order; concatenating the image feature map extracted by the convolutional neural network with the hidden state vector extracted by the recurrent neural network to form a fused feature representation; and inputting the fused feature into a fully connected neural network to output a result value representing the confidence level.

[0014] This preferred solution employs a convolutional neural network to extract spatial structural features from image-type observation data, combines this with the dynamic pattern modeling capability of a recurrent neural network for time-series data, and then concatenates the two types of features before inputting them into a fully connected neural network to output a unified confidence result. This enables the model to have a higher feature representation capability when processing both image and time-series information simultaneously, which helps improve the accuracy of confidence calculation. It is particularly suitable for confidence inference in multi-agent systems under multimodal information fusion scenarios.

[0015] As a preferred embodiment of the confidence-based multi-agent system interpretability enhancement method described in this invention, the following steps are included: constructing a visualization interface to display confidence levels, supporting the viewing of the confidence calculation basis and time changes of specific agent decisions; constructing an overall decision view, displaying the confidence status of multiple agents at the current time point, and using different colors to represent confidence intervals; constructing a time series view, plotting a confidence change curve with time step as the horizontal axis and confidence level as the vertical axis; constructing a relationship view, with nodes representing agents and edges representing the interaction relationships between agents, and adjusting the color and thickness of the edges according to the degree of interaction influence; and implementing interactive operations, including clicking to view the agent decision details at a specified time point, zooming to adjust the display ratio, and setting a confidence threshold for filtering and displaying.

[0016] This preferred solution constructs an overall decision view, a time series view, and a relationship view to achieve multi-dimensional display of confidence data. Combined with interactive methods such as clicking, zooming, and filtering, it enables users to intuitively view the confidence change trends and decision-making basis of different agents at different time points. This helps to enhance the understandability and traceability of the decision-making process and improve the application value of the system in debugging, analysis, and security review.

[0017] As a preferred embodiment of the confidence-based multi-agent system interpretability enhancement method described in this invention, the method of adjusting the strategy of the multi-agent algorithm according to the confidence of agent decisions, introducing confidence as a training objective during training, and optimizing the cooperation mechanism between agents using confidence information includes: setting a confidence threshold; if the confidence of a decision is lower than the confidence threshold, an exploratory strategy is adopted to select the optimal action under a different strategy with a certain probability; if the confidence of a decision is higher than the confidence threshold, an empirical strategy is adopted to directly select the optimal action under the current strategy; constructing a loss function that includes a confidence weighting term, giving higher loss weights to decision results with low confidence; and minimizing the loss function during model training, so that the model pays more attention to samples with low confidence during training.

[0018] This preferred scheme introduces a confidence threshold control training strategy exploration and utilization mechanism, combined with confidence weighted loss function design, so that the model can actively enhance its attention to low confidence decision samples during the training phase, thereby improving the model's generalization ability and policy robustness in uncertain environments, and improving the collaborative performance of multi-agent systems in highly dynamic or abnormal states.

[0019] This invention provides a confidence-based system for enhancing the interpretability of multi-agent systems.

[0020] To address the aforementioned technical problems, this invention provides the following technical solution: a confidence-based multi-agent system interpretability enhancement system, comprising: a data processing module, a confidence calculation module, a confidence correction module, a visualization module, and an algorithm optimization module; the data processing module is used to collect various information from each agent in the multi-agent system and perform preprocessing to obtain preprocessed information; the confidence calculation module is used to receive the preprocessed information based on a deep learning model and output a confidence value; the confidence correction module is used to correct the confidence value by combining environmental changes and information interaction between agents; the visualization module is used to construct a visualization interface to display the confidence, supporting the viewing of the basis for confidence calculation and time changes of specific agent decisions; the algorithm optimization module is used to adjust the strategy of the multi-agent algorithm according to the confidence of agent decisions, introduce confidence as a training target during training, and utilize confidence information to optimize the cooperation mechanism between agents.

[0021] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the confidence-based multi-agent system interpretability enhancement method.

[0022] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the aforementioned confidence-based multi-agent system interpretability enhancement method.

[0023] The beneficial effects of this invention are: It effectively improves the interpretability of the decision-making process in multi-agent systems, optimizes algorithm performance, and enhances the collaborative ability among agents. The method and system are universal and applicable to various multi-agent reinforcement learning algorithms, providing strong support for multi-agent applications in fields such as autonomous driving, smart grids, and drone swarms, and promoting the widespread application and development of multi-agent technology in complex tasks. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 The above is a flowchart of an overall method for enhancing the interpretability of a multi-agent system based on confidence, provided as an embodiment of the present invention. Detailed Implementation

[0026] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0027] Example 1, referring to Figure 1 This is one embodiment of the present invention, which provides a method for enhancing the interpretability of a multi-agent system based on confidence, comprising: S1. Collect various information from each agent in the multi-agent system and preprocess it to obtain preprocessed information.

[0028] S2. Receive preprocessed information based on a deep learning model and output confidence values.

[0029] S3. Adjust the confidence value by combining environmental changes and information interaction between agents.

[0030] S4. Construct a visual interface to display confidence levels, supporting the viewing of the basis for confidence level calculations and changes over time for specific agent decisions.

[0031] S5. Adjust the strategy of the multi-agent algorithm based on the confidence of the agent's decision. Introduce confidence as a training objective during the training process and use confidence information to optimize the cooperation mechanism between agents.

[0032] It should be noted that in multi-agent systems, the complex operating environments of each agent, the heterogeneity of input information, and the strong coupling of cooperative behaviors lead to poor interpretability of their behavior, making the visualization and control of confidence-based decisions quite challenging. On the one hand, due to the extensive interactions between agents, there is a high degree of uncertainty in the decision-making process, making it difficult for traditional fixed-strategy mechanisms to reflect the actual level of confidence in decisions. On the other hand, the lack of standardized visualization methods also limits the application of the system in security analysis, anomaly tracing, and deployment evaluation.

[0033] Therefore, to address the aforementioned issues, the system employs steps S1–S5. First, it collects the structured input information of the multi-agent system and performs preprocessing operations such as normalization to ensure the input data is standardized and consistent. Then, based on the deep learning model, it outputs the confidence value of each agent's current action, establishing a quantitative indicator of its decision-making credibility. When the environmental state changes or information interaction occurs between agents, the system dynamically corrects the existing confidence values, reflecting the real-time reliability of its decision-making process. Next, the confidence results are visualized through an overall decision view and time series changes, assisting external observers in understanding the agent's behavioral logic. Finally, the strategy structure of the multi-agent algorithm is adjusted according to the confidence level: a more exploratory strategy is used for low confidence, while an empirical strategy is used for high confidence. Confidence is also introduced as a loss control target during training to achieve more stable collaborative learning and stronger system interpretability. This accurately assesses the confidence of each agent's decision in the multi-agent system, quantifies the reliability of decisions, and provides users with clear decision-making basis. A confidence-based visualization interface is constructed to intuitively display the multi-agent decision-making process, helping researchers and users to gain a deeper understanding of the agents' behavioral patterns and policy learning processes. Confidence analysis is used to optimize multi-agent algorithms, improving algorithm performance, reducing training costs, and enhancing the algorithm's adaptability to different tasks and environments.

[0034] Example 2, an embodiment of the present invention, provides a method for enhancing the interpretability of a multi-agent system based on confidence, based on the previous embodiment, comprising: In this embodiment of the application, the information of each agent in step S1 may include observation data, action selection, reward feedback, and environmental state information, which are used to collect the observation data of each agent in the multi-agent system. Action selection Rewards and Feedback and environmental status information .

[0035] In one alternative implementation, the information of each agent may further include the agent's historical behavior trajectory and communication message log, wherein the behavior trajectory refers to the state transition record over a period of time, and the communication message log includes the information content and timestamps exchanged with other agents, which is used to enhance the ability to model collaborative intentions.

[0036] In another optional implementation, the information of each agent may also include the goal planning intention and constraint information, wherein the planning intention is the local task goal assigned by the scheduling module, and the constraint information includes physical restrictions, resource usage permissions, etc., which are used to assist the confidence calculation model in determining the reasonableness range of the action.

[0037] This invention effectively enhances the ability of confidence modeling to express the state of complex systems by introducing multi-dimensional input information, including perception, decision-making, feedback, and environment. This enables the model to have stronger adaptability and decision interpretation capabilities when dealing with multi-source heterogeneous inputs, thereby improving the overall system's understandability.

[0038] Furthermore, in step S1, information from each agent in the multi-agent system is collected and preprocessed to obtain preprocessed information, including the following steps A1-A3: A1. Collect various information during the execution process from multiple intelligent agents.

[0039] A2. Perform integrity checks and format standardization on the collected information.

[0040] A3. Preprocess the formatted information to obtain preprocessed information.

[0041] Specifically, in step A2, observation data from each agent in the multi-agent system are collected. Action selection Rewards and Feedback and environmental status information Then, it is cleaned and normalized.

[0042] Specifically, in step A3, for the observation data Normalization can be performed using the minimum-maximum normalization method, with the following formula: in, and These are the minimum and maximum values ​​in the observed dataset, respectively. For action selection data, preprocessing methods such as one-hot encoding based on the action space can be used to transform it into a form suitable for model input. Assuming the action space size is m, and the actions... The corresponding one-hot encoding vector is Only the index position corresponding to the action is 1, and the rest are 0.

[0043] In this embodiment of the application, the deep learning model in step S2 can be a combination of convolutional neural network and recurrent neural network, wherein the convolutional neural network is used to extract spatial features of image observation data, the recurrent neural network is used to process time series information, and the features output by the two are concatenated and input into a fully connected neural network for confidence value calculation.

[0044] In an alternative implementation, the deep learning model can also be a combination of a gated recurrent unit network and an attention mechanism module, wherein the gated recurrent unit network is used to capture short-term and long-term dependencies in the sequence data, and the attention mechanism is used to dynamically adjust the weight distribution of information at each time step in the confidence output to enhance the model's ability to focus on key moments.

[0045] In another alternative implementation, the deep learning model can also be a superposition structure of graph neural network and feedforward neural network, wherein the graph neural network is used to model the structured interaction relationship between agents, and the feedforward neural network is used to map the fused graph features into confidence output, which is suitable for multi-agent task scenarios with strong structural dependencies.

[0046] This invention introduces a deep fusion structure of convolutional neural networks and recurrent neural networks, which enables the collaborative expression of image-based observation information and time-series information. It can accurately capture the complex mapping relationship between state and behavior in multi-agent systems, thereby improving the accuracy and stability of confidence calculation and enhancing the system's ability to identify uncertain behaviors.

[0047] Furthermore, in step S2, the preprocessed information is received based on the deep learning model, and the confidence value is output, including the following steps B1-B3: B1. The preprocessed information is fed into the deep learning model substructures used to process image information and sequence information, respectively.

[0048] B2. After extracting image features and temporal features, they are fused into a set of joint feature vectors.

[0049] B3. Input the joint feature vector into the fully connected network structure to generate confidence output values.

[0050] Furthermore, in step B1, the preprocessed information is fed into the deep learning model substructures used for processing image information and sequence information, respectively, including the following steps B11-B14: B11. Convolutional neural networks are used to process image-type observation data. They sequentially perform multiple convolution and pooling operations on the input image to generate image feature maps.

[0051] B12. Recurrent neural networks are used to process time series observation data and extract hidden state information at each time step in chronological order.

[0052] B13. The image feature map extracted by the convolutional neural network is concatenated with the hidden state vector extracted by the recurrent neural network to form a fused feature representation.

[0053] B14. Input the fused features into a fully connected neural network and output the result value representing the confidence level.

[0054] Specifically, in step B11, for image-type observation data, it is assumed that the observation data has undergone a series of convolutional and pooling layer operations of a convolutional neural network.

[0055] For the input image observation data After the convolution operation of the convolutional layer. Assume the l-th convolutional layer has... convolution kernels If the size of each convolutional kernel is k×k, then the j-th feature map of this layer... The calculation is as follows: Where * denotes a convolution operation. It is the i-th feature map of the previous layer. It's a bias. It is an activation function, such as the ReLU function.

[0056] Specifically, in step B12, for time series data This is processed using a recurrent neural network (such as LSTM). The computation process of an LSTM unit at each time step is as follows: Forgotten Gate : Input gate : Candidate memory units : memory unit : Output gate : Hidden state : in, These are the weight matrices for the forget gate, input gate, candidate memory units, and output gate, respectively. It is the corresponding bias term. This indicates element-wise multiplication. It is the hidden state from the previous moment. This is the input at the current moment. Finally, the feature map F processed by the CNN and the hidden state processed by the RNN are compared. The input is then fused and fed into a fully connected layer for confidence calculation, yielding the confidence value P for the decision.

[0057] Specifically, in step B14, for image-type observation data, the feature map F is obtained through the CNN processing flow as shown in the above formula; for time-series data, the hidden state is obtained through the RNN (such as LSTM) processing flow as shown in the above formula. The concatenation of F and h is then fed into the fully connected layer. The output of the fully connected layer is used to obtain the confidence value through an activation function (such as the softmax function). Assume the weight matrix of the fully connected layer is... , bias is The concatenated input vector is The confidence score P is calculated as follows: Furthermore, in step S3, the confidence value is corrected by combining environmental changes and information interaction between agents, including the following steps C1-C4: C1. Obtain information on current environmental changes and information exchange data between the agent.

[0058] C2. Determine whether environmental changes may affect the confidence level of the decision.

[0059] C3. Based on environmental change information and interaction data, call the preset confidence correction function to calculate the confidence correction value.

[0060] C4. Apply the correction value to the confidence output value to generate the corrected confidence result.

[0061] In this embodiment of the application, the determination of whether environmental changes may affect the decision confidence level in step C2 may include comparing the difference between the current environmental state and the historical environmental average. When the magnitude of the environmental state change exceeds a set threshold, it is determined that it may affect the current confidence level, and a confidence level correction process is triggered. For example, when changes in environmental temperature, humidity, traffic density, etc., are detected to exceed the system's set threshold range, the current state is considered abnormal and the confidence adjustment process is initiated.

[0062] In one alternative implementation, determining whether environmental changes may affect decision confidence may further include scoring the current environmental state based on a pre-trained environmental impact assessment model. If the assessment score is lower than the reliability threshold, the environmental changes are considered significant and there is a potential interference with the agent's decision-making. The original confidence level needs to be reduced to reflect decision uncertainty.

[0063] In another alternative implementation, determining whether environmental changes may affect decision confidence may also include calculating the similarity between the agent's current perception state and the initial state of task planning. If the similarity is lower than a set standard, it is determined that the current environment deviates from the task expectation, which may cause perception errors or behavioral deviations, thereby affecting the credibility of the confidence.

[0064] This invention achieves rapid identification of external conditions affecting confidence by using an environmental state change and threshold judgment mechanism. This enables the system to dynamically adjust the confidence level when environmental anomalies or disturbances occur, avoiding misjudgments due to excessively high confidence caused by drastic environmental changes, and enhancing the robustness and decision credibility of the system in dynamic environments.

[0065] In this embodiment of the application, calculating the confidence correction value in step C3 may include calling a confidence correction function, taking current environmental change information and information interaction data between intelligent agents as input, outputting a correction scaling factor, and applying this factor to the original confidence value to obtain the corrected confidence value. For example, in an intelligent driving scenario, when special conditions such as road icing are detected, the correction function will output a factor lower than 1, thereby reducing the original confidence value.

[0066] In an optional implementation, calculating the confidence correction value may further include: constructing a state difference vector based on the difference between the current environmental state and the historical stable environmental state, and inputting the vector into an empirical model to calculate the correction value. This model is trained based on historical data and is used to quantify the degree of influence of different environmental changes on the confidence level.

[0067] In another optional implementation, calculating the confidence correction value may also include: using matching templates in the rule template library, calling the corresponding correction rules according to the type of environmental change event (such as sudden weather changes, equipment interference, communication packet loss, etc.), and directly outputting the correction value for rapid response to high-risk change events.

[0068] This invention establishes a confidence correction function to handle the uncertainty of environmental changes and information interaction on the agent's decision-making in a structured manner, thereby realizing the dynamic correction capability of the confidence output. This makes the final confidence result more realistically reflect the current environment of the system and improves the adaptability and reliability of confidence assessment in complex scenarios.

[0069] Specifically, in step C4, the confidence level output by the confidence calculation module is corrected based on environmental changes and information interaction between agents. Let the environmental change information be... The information exchange data between intelligent agents is The confidence correction function is The corrected confidence level for: For example, when significant environmental changes occur, such as detecting road icing in an intelligent driving scenario, environmental change information... This can be represented as a vector containing features such as road conditions and weather. If, based on experience, icy roads increase the uncertainty of an agent's decision-making, a design can be developed... The function reduces the confidence level by a certain percentage, such as ,in This is a correction factor determined based on the degree of environmental change, and .

[0070] Furthermore, in step S4, a visualization interface for displaying confidence levels is constructed, supporting the viewing of the basis for confidence level calculations and changes over time for specific agent decisions, including the following steps D1-D4: D1. Construct an overall decision view that displays the confidence status of multiple agents at the current time point and uses different colors to represent confidence intervals.

[0071] D2. Construct a time series view and plot the confidence curve with time step as the horizontal axis and confidence level as the vertical axis.

[0072] D3. Construct a relationship view, where nodes represent agents and edges represent the interaction relationships between agents, and adjust the color and thickness of the edges according to the degree of interaction impact.

[0073] D4. Implement interactive operations, including clicking to view the agent's decision details at a specified time point, zooming to adjust the display ratio, and setting confidence thresholds for filtering and display.

[0074] Specifically, in step D1, different colors are used to represent different confidence intervals in the overall decision view. Assume the confidence interval is divided into [0, 0.4) as red, [0.4, 0.7) as yellow, and [0.7, 1] as green. For the decision of agent j at a certain moment, if its confidence level... If the value is in the interval [0.4, 0.7), then the decision identifier of the agent will be displayed in yellow in the view.

[0075] In the overall decision view, when a user clicks on a decision icon of an agent, the system displays detailed information about that agent's decision, including the decision action 'a', the observed data 'o', the original confidence level 'P', and the adjusted confidence level. Simultaneously, it demonstrates the key regions of the CNN feature map and the key time steps of the RNN hidden states during the confidence calculation process. Let the CNN feature map be F, and the importance weights of the feature map are obtained through methods such as Gradient Weighted Class Activation Mapping (Grad-CAM). The feature map of the key region It can be represented as: For the hidden state h of the RNN, the gradient of the hidden state with respect to the confidence is calculated. Find the time step with the larger absolute value of the gradient as the critical time step.

[0076] Specifically, in step D2, a line graph is plotted in the time series view, with time step as the x-axis and confidence level as the y-axis. For agent k, its confidence level at different time steps t is... Points forming a line chart .

[0077] In a time series view, when the user zooms in or out, let the zoom level be 's'. To enable zooming in, the number of time steps displayed on the timeline is reduced, the display area is narrowed, and the confidence scale on the vertical axis is adjusted accordingly to more clearly show details. To reduce the scope of the operation, the number of time steps displayed on the timeline is increased, and the display range is widened. Assuming the original timeline range is... The scaled timeline range becomes .

[0078] Specifically, in the relationship view of step D3, nodes represent agents, and edges represent the influence relationships between agents. The color intensity or thickness of the edges can be determined based on the degree of influence of the decision confidence between agents. Assume that the degree of influence of agent i on the decision confidence of agent j is... By setting thresholds and mapping relationships, Transform it into a visual attribute of the edge.

[0079] Specifically, in step D4, the interactive functions of the visualization interface include viewing detailed decision information and confidence calculation basis for a specific agent or time point through mouse clicks and zoom operations, as well as setting filter conditions to view the decision-making situation of agents that meet specific conditions. When a user clicks on the decision marker of an agent at a specific time point on the visualization interface, the system will obtain the decision action of agent j at that time point t based on the click event. Observational data Regarding the basis for confidence calculation, assuming the confidence calculation model is M, it can be presented by displaying the outputs of intermediate layers in model M that are relevant to the agent's decision-making, such as feature maps of convolutional layers and input / output information of fully connected layers. For example, displaying the feature map of convolutional layer l. and the input vector of the fully connected layer and output confidence value The calculation process: in, and These are the weights and biases of the fully connected layer. During scaling, the system adjusts the size and display range of the visualization elements according to the scaling ratio 's'. For example, for a line chart in a time series view, if the scaling ratio is 's', the number of time steps displayed on the time axis becomes 's' times the original number, and the vertical axis range of the line chart is adjusted accordingly to more clearly show the trend of confidence over time. When setting filtering conditions, it is assumed that the filtering condition is that the agent's decision confidence is greater than a threshold. The system will iterate through the decision data of all agents and filter out those that meet the requirements. The decision data is presented in a visualization interface, including highlighting the decision identifiers of these agents in the overall decision view and plotting only the confidence line graphs of these agents in the time series view.

[0080] Furthermore, in step S5, the strategy of the multi-agent algorithm is adjusted based on the confidence level of the agent's decision. Confidence level is introduced as a training objective during the training process, and the cooperation mechanism between agents is optimized using confidence level information, including the following steps E1-E4: E1. Set a confidence threshold. If the confidence of a decision is lower than the confidence threshold, an exploratory strategy is adopted to select the optimal action under a different strategy with a certain probability.

[0081] E2. If the confidence level of a decision is higher than the confidence threshold, then an empirical strategy is adopted, and the optimal action under the current strategy is directly selected.

[0082] E3. Construct a loss function that includes a confidence weighting term, and give higher loss weights to decision results with low confidence.

[0083] E4. Minimize the loss function during model training to make the model pay more attention to samples with low confidence during training.

[0084] Specifically, in steps E1 and E2, a more exploratory strategy is adopted when the agent's decision confidence is low, and a strategy that utilizes existing experience is adopted when the confidence is high. A confidence threshold is set. The confidence level of the agent at a certain decision moment At that time, take Exploratory strategies. For example, using a greedy strategy to select actions based on probability. Randomly select actions, with probability Choose the optimal action under the current strategy. It can be dynamically adjusted according to the confidence level, such as .when When necessary, a strategy that leverages existing experience is adopted, directly selecting the optimal action under the current strategy. ,Right now: in, It is the action value function, and s is the current state.

[0085] When the agent's decision confidence is low, such as ( To set a threshold, a more exploratory strategy can be adopted. For example, in policy gradient-based algorithms, the parameter update method of the policy network can be adjusted. Let the parameters of the policy network be... The original policy gradient update formula is: in, For learning rate, To take an action in state s The probability, It is the action value; when the confidence level is low, an exploration coefficient can be introduced. To adjust the update direction, the new update formula becomes: This enhances the exploratory aspect.

[0086] When the confidence level is high, such as The strategy utilizes existing experience, directly selecting the optimal action based on the current policy network, i.e.: Specifically, in step E3, during training, the confidence-related loss function is optimized to increase the weight of decisions with lower confidence during training, enabling the algorithm to learn more reliable decision-making strategies. Let the original loss function be... Add confidence-related loss terms Afterwards, the total loss function L is: in, This is a balancing coefficient used to adjust the relative importance of the two loss terms. For decisions with lower confidence, their weight is increased. Assume that the confidence level of agent i's decision at time step t is... Defineable as follows: in, This is the original loss value of agent i at time step t. By minimizing the total loss function L, during training, the original loss value corresponding to decisions with low confidence is... Their larger contribution to the total loss prompts the algorithm to pay more attention to these decisions, thereby learning more reliable decision-making strategies.

[0087] To improve the cooperation mechanism between agents, the information sharing weights between agents are adjusted based on confidence levels. Assuming that agents i and j share information, the weights for information sharing are... They can be determined based on their confidence level. For example, Then, during information fusion, agent i receives information from agent j. And update its own information The formula is: Furthermore, at low confidence levels, in addition to adjusting the policy gradient update formula, a Boltzmann exploration policy can also be adopted. Let the probability of choosing action a in state s be... According to the Boltzmann distribution: in, It is a temperature parameter; it increases when the confidence level is low. The value of makes action selection more random and enhances the exploratory nature of the activity.

[0088] At high confidence levels, the empirical policy can update the current policy based on the historical best policy. Let the historical best policy be... The current strategy is The updated formula is: in, It is the update coefficient. .

[0089] In terms of optimizing the loss function during training, in addition to the method of introducing a confidence-related loss term mentioned above, adaptive adjustment of the balancing coefficient can also be used. The method is as follows. Let the rate of change of the loss function L with respect to the number of training steps N be... ,when A larger value indicates unstable training; the value should be increased appropriately. To focus more on confidence-related losses; when A smaller value indicates more stable training; therefore, it's advisable to reduce the value appropriately. To focus more on the original loss function.

[0090] Example 3 is an embodiment of the present invention. This embodiment provides a confidence-based multi-agent system interpretability enhancement system, including a data processing module, a confidence calculation module, a confidence correction module, a visualization module, and an algorithm optimization module.

[0091] The data processing module is used to collect various information from each agent in the multi-agent system and perform preprocessing to obtain preprocessed information.

[0092] The confidence calculation module is used to receive preprocessed information based on a deep learning model and output confidence values.

[0093] The confidence correction module is used to correct the confidence value by combining environmental changes and information interaction between agents.

[0094] The visualization module is used to build a visual interface to display confidence levels, allowing users to view the basis for confidence level calculations and changes over time for specific agent decisions.

[0095] The algorithm optimization module is used to adjust the strategy of the multi-agent algorithm based on the confidence of the agent's decision. During the training process, confidence is introduced as a training objective, and confidence information is used to optimize the cooperation mechanism between agents.

[0096] This embodiment also provides an electronic device applicable to a confidence-based multi-agent system interpretability enhancement method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the confidence-based multi-agent system interpretability enhancement method proposed in the above embodiment.

[0097] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a confidence-based method for enhancing the interpretability of a multi-agent system as proposed in the above embodiments.

[0098] The storage medium proposed in this embodiment and the method for enhancing the interpretability of a confidence-based multi-agent system proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0099] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0100] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for enhancing the interpretability of a multi-agent system based on confidence, characterized in that: include, Collect various information from each agent in the multi-agent system and preprocess it to obtain preprocessed information; The system receives preprocessed information based on a deep learning model and outputs a confidence score. The confidence value is adjusted by combining environmental changes and information interaction between intelligent agents; Build a visual interface to display confidence levels, supporting the viewing of the basis for confidence level calculations and changes over time for specific agent decisions; The strategy of the multi-agent algorithm is adjusted based on the confidence level of the agent's decision. Confidence level is introduced as a training objective during the training process, and the cooperation mechanism between agents is optimized using confidence level information.

2. The method for enhancing the interpretability of a multi-agent system based on confidence as described in claim 1, characterized in that: The process involves collecting various information from each agent in the multi-agent system and preprocessing it to obtain preprocessed information, including: Collect various information during the execution process from multiple intelligent agents; Perform integrity checks and format standardization on the collected information; The formatted information is preprocessed to obtain preprocessed information.

3. The method for enhancing the interpretability of a multi-agent system based on confidence as described in claim 2, characterized in that: The process of receiving preprocessed information based on a deep learning model and outputting confidence values ​​includes... The preprocessed information is fed into the deep learning model substructures used to process image information and sequence information, respectively. After extracting image features and temporal features, they are fused into a joint feature vector. The joint feature vector is input into a fully connected network structure to generate confidence output values.

4. The method for enhancing the interpretability of a multi-agent system based on confidence as described in claim 3, characterized in that: The process of adjusting the confidence value by combining environmental changes and information interaction between intelligent agents includes, Acquire information about current environmental changes and data exchanged between intelligent agents; Determine whether environmental changes are likely to affect decision confidence levels; Based on environmental change information and interaction data, a preset confidence correction function is invoked to calculate the confidence correction value; The correction value is applied to the confidence output value to generate the corrected confidence result.

5. The method for enhancing the interpretability of a multi-agent system based on confidence as described in claim 3, characterized in that: The step of feeding the preprocessed information into the deep learning model substructures for processing image and sequence information includes, Convolutional neural networks are used to process image-like observation data. They sequentially perform multiple convolution and pooling operations on the input image to generate image feature maps. Recurrent neural networks are used to process time series observation data and extract hidden state information at each time step in chronological order. The image feature map extracted by the convolutional neural network is concatenated with the hidden state vector extracted by the recurrent neural network to form a fused feature representation; The fused features are input into a fully connected neural network, and the output is a result value representing the confidence level.

6. The method for enhancing the interpretability of a multi-agent system based on confidence as described in claim 5, characterized in that: The constructed visualization interface displays confidence levels, supporting the viewing of the basis for confidence level calculations and their changes over time for specific agent decisions, including: Construct an overall decision view that displays the confidence status of multiple agents at the current time point and uses different colors to represent confidence intervals; Construct a time series view and plot the confidence level change curve with time step as the horizontal axis and confidence level as the vertical axis; Construct a relationship view, where nodes represent agents and edges represent the interaction relationships between agents, and adjust the color and thickness of the edges according to the degree of interaction impact; Enables interactive operations, including clicking to view the agent's decision details at a specified time point, zooming to adjust the display ratio, and setting confidence thresholds for filtering and display.

7. The method for enhancing the interpretability of a multi-agent system based on confidence as described in claim 6, characterized in that: The strategy of adjusting the multi-agent algorithm based on the confidence level of the agent's decisions includes introducing confidence level as a training objective during training and utilizing confidence information to optimize the cooperation mechanism between agents. Set a confidence threshold. If the confidence of a decision is lower than the confidence threshold, an exploratory strategy is adopted to select the optimal action that is not the current strategy with a certain probability. If the confidence level of a decision is higher than the confidence threshold, then an empirical strategy is adopted, and the optimal action under the current strategy is directly selected. Construct a loss function that includes a confidence weighting term, and give higher loss weights to decision results with low confidence. Minimize the loss function during model training to make the model pay more attention to samples with low confidence during training.

8. A confidence-based multi-agent system interpretability enhancement system, employing the confidence-based multi-agent system interpretability enhancement method as described in any one of claims 1 to 7, characterized in that, include: Data processing module, confidence calculation module, confidence correction module, visualization module, and algorithm optimization module; The data processing module is used to collect various information from each agent in the multi-agent system and perform preprocessing to obtain preprocessed information. The confidence calculation module is used to receive preprocessed information based on a deep learning model and output a confidence value. The confidence correction module is used to correct the confidence value by combining environmental changes and information interaction between intelligent agents; The visualization module is used to build a visualization interface to display confidence levels, and supports viewing the basis for confidence level calculation and time changes of a specific agent's decision. The algorithm optimization module is used to adjust the strategy of the multi-agent algorithm based on the confidence level of the agent's decision. During the training process, confidence level is introduced as the training objective, and confidence level information is used to optimize the cooperation mechanism between agents.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the confidence-based multi-agent system interpretability enhancement method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the confidence-based multi-agent system interpretability enhancement method according to any one of claims 1 to 7.

Images