Multi-agent dynamic weighted decision-making method based on self continuity

Abstract

The application provides a multi-agent dynamic weighting decision-making method based on self continuity, and relates to the technical field of deep learning model training, and comprises the following steps: S1, collecting and storing an initial state vector of each agent; S2, periodically collecting a current state of each agent; S3, calculating a self continuity score of each agent; S4, obtaining a voting value of each agent for a current decision theme; S5, calculating a final decision weight of each agent; S6, calculating a final decision score of the system; and S7, outputting a final decision instruction. By introducing the self continuity score, quantitative evaluation of the health state of each agent is realized, the final decision weight of the agent is dynamically adjusted according to the self continuity score of the agent, error views of one or more abnormal agents are prevented from polluting the final decision result, and the robustness and stability of the system are improved.

Description

Technical Field

[0001] This invention relates to the field of deep learning model training technology, and in particular to a multi-agent dynamic weighted decision-making method based on self-continuity. Background Technology

[0002] In existing multi-agent decision-making systems, the system typically consists of multiple agents with different roles and functions. Each agent outputs its decision-making viewpoint based on its own perceptions and policies, and the system aggregates these viewpoints to form a final decision. However, existing technologies have the following shortcomings in decision aggregation mechanisms and agent state monitoring: 1. Traditional multi-agent decision-making aggregation methods often employ majority voting or fixed-weighting. Majority voting only counts the number of opinions output by each agent, failing to differentiate the quality of those opinions. While fixed-weighting considers the differences in agent roles, the weights remain unchanged once set. Both methods suffer from a fatal flaw: they cannot identify "crazy" agents. A "crazy" agent refers to one whose decision-making logic is confused, whose values ​​deviate, and whose output is unstable due to contextual interference, adversarial attacks, internal logic drift, or training data contamination. Under traditional mechanisms, an agent that has already experienced value drift will still have the same opinion weight as a stable agent, causing its erroneous opinions to contaminate the final decision result and even triggering system-level decision failure.

[0003] 2. Existing technologies cannot dynamically adjust the decision-making influence of agents based on their real-time status. Even if an agent makes consecutive mistakes or exhibits obvious logical disorder, the system cannot reduce its weight or limit its speaking power in real time. This static decision aggregation mechanism makes the system lack adaptive purification capabilities when facing agent performance degradation or malicious attacks, making it difficult to guarantee the stability and reliability of long-term operation.

[0004] 3. Current security protection for multi-agent systems is mostly done through post-event correction, such as identifying abnormal agents through log analysis and manual intervention. This approach is slow to respond and cannot provide early warnings when an agent "goes crazy" or intercept its harmful decisions. Once the abnormal agent's opinion is adopted, it will cause irreversible losses.

[0005] 4. In the training process of deep learning models, traditional loss functions (such as cross-entropy loss and mean squared error) only focus on the completion of task objectives (such as classification accuracy and prediction error), and do not consider whether the model has undergone "intrinsic value drift" during training. That is, whether the model's internal representation still retains its initial "self", which may cause the model to perform well in task indicators, but its decision logic and value orientation have deviated from the initial design and become unexplainable.

[0006] To address this, a multi-agent dynamic weighted decision-making method based on self-continuity is proposed. Summary of the Invention

[0007] In view of this, the present invention provides a multi-agent dynamic weighted decision-making method based on self-continuity to solve or alleviate one of the technical problems existing in the prior art, and at least provides a beneficial option.

[0008] The technical solution of this invention is implemented as follows: a multi-agent dynamic weighted decision-making method based on self-continuity includes the following steps: Step S1: During the system initialization phase, collect and store the initial state vector of each agent. The initial state vector is generated by a self-representation encoder. It is obtained by encoding the initial parameters, value anchors, or behavioral patterns of the intelligent agent; Step S2: During system operation, periodically collect the current state of each agent. And through the self-representation encoder Map it to the current state vector ; Step S3: Calculate the ego continuity score for each agent. The self-continuity score is used to quantify the degree of deviation of the agent's current state from its initial state, and is calculated using the following formula:

[0009] in, The square of the Euclidean distance. The value range is [0, 1]; Step S4: Obtain the voting value Vi for each agent regarding the current decision topic. The voting value Vi can be a discrete or continuous value, used to express the agent's preference for a certain action plan. Step S5: Based on the base weights of each agent Based on the self-continuity score continuityi, calculate the final decision weight for each agent. The calculation formula is:

[0010] in, Determined by the pre-defined role of the agent; Step S6: Calculate the weighted average of all agent votes to obtain the final decision score F of the system. The calculation formula is as follows:

[0011] Where n is the number of agents involved in the decision-making process; Step S7: Compare the final decision score F with the preset decision threshold, and output the final decision instruction.

[0012] A further preferred embodiment includes a state pre-check step before calculating the self-continuity score in step S3. Before an agent is about to make a decision, the system uses the agent's current state as a basis. Given the decision to be output, simulate the possible future states that may result from the execution of the decision, and calculate the expected self-continuity score. If the self-continuity score is lower than the first preset threshold, the agent's current decision output is rejected, and a self-check process is triggered, requiring the agent to regenerate the decision. If the self-continuity score is lower than the second preset threshold and the second preset threshold is lower than the first preset threshold, then the agent's state is forcibly rolled back to the initial state, and its decision-making power is suspended until recalibration is completed.

[0013] A further preferred embodiment includes a three-level safety line fuse mechanism, wherein the three-level safety line fuse mechanism is based on deviation. Trigger different levels of security responses;

[0014] When deviation i ≥ Yellow line threshold and deviation i When the threshold value is less than the red line threshold, the system logs the event, reduces the final decision weight of the agent in subsequent decisions, and alerts the system administrator. The yellow line threshold value ranges from [0.2, 0.4]. When deviation i ≥ Red line threshold and deviation i When the deviation falls below the black line threshold, the system rejects the agent's decision, triggers a self-check process, requires the agent to regenerate its opinion, and temporarily freezes the agent's voting rights until its deviation drops below the red line threshold, which ranges from [0.4, 0.6]. When deviation i When the value is greater than or equal to the black line threshold, the system will force the agent's state to roll back to the initial state, suspend its speaking rights until stability is restored, and trigger the full system audit process. The value range of the black line threshold is [0.7, 0.9].

[0015] A further preferred embodiment: The self-test process includes: The agent enters an isolated sandbox environment, replays its recent decision-making path in the sandbox, and analyzes the key decision nodes that lead to increased deviation. The system generates a difference report by comparing the current decision path with the initial decision path, and attempts to repair the agent by fine-tuning parameters or reverting to the checkpoint before the deviation. If the self-check process successfully reduces the deviation to below the red line threshold, the agent's normal decision-making power is restored; otherwise, the frozen state is maintained and manual intervention is awaited.

[0016] Further preferred: the self-representation encoder For a trainable neural network model, an ego continuity loss function is introduced during its training process. The self-continuity loss function is defined as follows:

[0017] During encoder training, the total loss function is:

[0018] in, For the task loss function, This is a balance coefficient used to control the strength of the self-continuity constraint; it is achieved by minimizing... This allows the encoder to maintain sensitivity to changes in the agent's state while completing a specific task, reflecting the degree of drift in its intrinsic value.

[0019] Further preferred: The basic weight is determined by the role type of the agent, which includes, but is not limited to, risk-sensitive agents, value-investing agents, trend-following agents, and arbitrage agents.

[0020] Further preferred embodiments include, but are not limited to, binary discrete values, multivariate discrete values, continuous values, and probability values.

[0021] Further optimization: While outputting the final decision result, the system also includes the final decision weight, self-continuity score, voting value, and deviation level of each agent, forming a visualized decision tracing report.

[0022] In a second aspect, this aspect provides an electronic device comprising at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform a multi-agent dynamic weighted decision-making method based on self-continuity.

[0023] Thirdly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute a multi-agent dynamic weighted decision-making method based on self-continuity.

[0024] The embodiments of the present invention have the following advantages due to the adoption of the above technical solutions: I. This invention achieves a quantitative assessment of the health status of each agent by introducing a self-continuity score. During decision aggregation, the system dynamically adjusts the final decision weight based on the agent's self-continuity score, enabling the system to automatically identify and suppress the negative impact of "crazy" agents, preventing the erroneous views of one or more abnormal agents from contaminating the final decision result, and improving the robustness and stability of the system.

[0025] Second, the self-continuity loss function proposed in this invention simultaneously constrains the task objective and self-consistency during model training. By introducing the self-continuity loss as a regularization term into the total loss function, the model maintains its intrinsic consistency with its initial self while pursuing optimal task performance, preventing irreversible value drift during training and ensuring the controllability of the model's decision-making logic.

[0026] Third, while outputting the final decision result, this invention automatically appends the final decision weight, self-continuity score, voting value, and deviation level for each agent, forming a visualized decision tracing report. The report records the health status of each agent and its contribution to the final decision, enabling system administrators to intuitively understand the decision-making process, quickly identify potentially risky agents, optimize system parameter configurations, and reduce the complexity of system operation and maintenance.

[0027] The above overview is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of the invention will become readily apparent from the accompanying drawings and the following detailed description. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0030] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0031] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0032] like Figure 1 As shown, this embodiment of the invention provides a multi-agent dynamic weighted decision-making method based on self-continuity, characterized by the following steps: Step S1: During the system initialization phase, collect and store the initial state vector of each agent. The initial state vector is generated by a self-representation encoder. It is obtained by encoding the initial parameters, value anchors, or behavioral patterns of the intelligent agent; Step S2: During system operation, periodically collect the current state of each agent. And through the self-representation encoder Map it to the current state vector ; Step S3: Calculate the ego continuity score for each agent. The self-continuity score is used to quantify the degree of deviation of the agent's current state from its initial state, and is calculated using the following formula:

[0033] in, The square of the Euclidean distance. The value range is [0, 1]; Step S4: Obtain the voting value Vi for each agent regarding the current decision topic. The voting value Vi can be a discrete or continuous value, used to express the agent's preference for a certain action plan. Step S5: Based on the base weights of each agent Based on the self-continuity score continuityi, calculate the final decision weight for each agent. The calculation formula is:

[0034] in, Determined by the pre-defined role of the agent; Step S6: Calculate the weighted average of all agent votes to obtain the final decision score F of the system. The calculation formula is as follows:

[0035] Where n is the number of agents involved in the decision-making process; Step S7: Compare the final decision score F with the preset decision threshold, and output the final decision instruction.

[0036] Steps S1 to S7 constitute the core method of this invention. The initial state vector is equivalent to the anchor point of the agent, which is mapped to a high-dimensional space by the encoder to facilitate subsequent quantification of deviations. Step S2 updates the current state in real time, step S3 calculates the continuity score (the closer the value is to 1, the more stable it is), step S4 collects the votes of each agent, step S5 dynamically adjusts the weights using the continuity score, step S6 calculates the weighted average to obtain the comprehensive decision score, and step S7 outputs the final instruction based on the threshold. The entire process realizes the perception of the agent's health status and the dynamic adjustment of the influence of decision-making.

[0037] In this embodiment, specifically, before calculating the self-continuity score in step S3, a state pre-check step is also included: Before an agent is about to make a decision, the system uses the agent's current state as a basis. Given the decision to be output, simulate the possible future states that may result from the execution of the decision, and calculate the expected self-continuity score. If the self-continuity score is lower than the first preset threshold, the agent's current decision output is rejected, and a self-check process is triggered, requiring the agent to regenerate the decision. If the self-continuity score is lower than the second preset threshold and the second preset threshold is lower than the first preset threshold, then the agent's state is forcibly rolled back to the initial state, and its decision-making power is suspended until recalibration is completed.

[0038] The state pre-check mechanism is a preventative measure. Before the agent makes a formal statement, the system simulates the execution of the decision, predicts the potential state changes it might trigger, and calculates the expected continuity score. If the expected score is too low (below a first threshold), it indicates that the decision might cause the agent to "go crazy," so it is intercepted in advance; if it is even lower (below a second threshold), it indicates that a serious drift is about to occur, and the system is directly rolled back to the initial state. This mechanism moves the safety checkpoint forward to before the decision is made, preventing erroneous viewpoints from contaminating the system.

[0039] In this embodiment, specifically, a three-level safety line fuse mechanism is included, which is based on deviation. Trigger different levels of security responses;

[0040] When deviation i ≥ Yellow line threshold and deviation i When the threshold value is less than the red line threshold, the system logs the event, reduces the final decision weight of the agent in subsequent decisions, and alerts the system administrator. The yellow line threshold value ranges from [0.2, 0.4]. When deviation i ≥ Red line threshold and deviation iWhen the deviation falls below the black line threshold, the system rejects the agent's decision, triggers a self-check process, requires the agent to regenerate its opinion, and temporarily freezes the agent's voting rights until its deviation drops below the red line threshold, which ranges from [0.4, 0.6]. When deviation i When the value is greater than or equal to the black line threshold, the system will force the agent's state to roll back to the initial state, suspend its speaking rights until stability is restored, and trigger the full system audit process. The value range of the black line threshold is [0.7, 0.9].

[0041] The three-tiered safety line circuit breaker mechanism divides deviation into four intervals: normal, yellow line warning, red line interception, and black line rollback. When a yellow line warning is triggered, the system automatically reduces the agent's influence and alerts the administrator. When a red line interception is triggered, the system not only rejects the current decision but also triggers a self-check process to attempt repair. Black line rollback is the strongest measure, forcibly restoring the initial state and suspending permissions. These three lines of defense are progressively stronger, ensuring the system can respond promptly when an agent malfunctions, preventing a single point of failure from causing a global crash.

[0042] In this embodiment, the self-test process specifically includes: The agent enters an isolated sandbox environment, replays its recent decision-making path in the sandbox, and analyzes the key decision nodes that lead to increased deviation. The system generates a difference report by comparing the current decision path with the initial decision path, and attempts to repair the agent by fine-tuning parameters or reverting to the checkpoint before the deviation. If the self-check process successfully reduces the deviation to below the red line threshold, the agent's normal decision-making power is restored; otherwise, the frozen state is maintained and manual intervention is awaited.

[0043] The self-check process is the agent's self-repair mechanism. When the agent's deviation reaches a red line, it is placed in a sandbox isolation environment to replay its recent decision-making history and locate the key nodes that caused the deviation. The system analyzes the cause of the drift by comparing the current path with the initial path, and then attempts to fine-tune parameters or roll back to a previous stable checkpoint. If the repair is successful, the agent returns to normal; otherwise, manual intervention is required, thus achieving automated problem diagnosis and repair and reducing manual maintenance costs.

[0044] In this embodiment, specifically, the self-representation encoder For a trainable neural network model, an ego continuity loss function is introduced during its training process. The self-continuity loss function is defined as follows:

[0045] During encoder training, the total loss function is:

[0046] in, For the task loss function, This is a balance coefficient used to control the strength of the self-continuity constraint; it is achieved by minimizing... This allows the encoder to maintain sensitivity to changes in the agent's state while completing a specific task, reflecting the degree of drift in its intrinsic value.

[0047] encoder The training employs a self-continuity loss function. Traditional task losses (such as classification losses) only concern task performance, which can easily lead to representational drift within the model. This invention introduces... By constraining the distance between the current state and the initial state in the feature space, the model is forced to optimize the task objective without deviating from its initial self. λ adjusts the weights of both, enabling the model to perform the task well while maintaining the controllability of state changes.

[0048] In this embodiment, specifically, the basic weights are determined by the agent's role type, which includes, but is not limited to, risk-sensitive agents, value-investing agents, trend-following agents, and arbitrage agents. The basic weights reflect the inherent importance of different roles in decision-making.

[0049] In this embodiment, the voting value can be expressed in various forms, including but not limited to binary discrete values, multivariate discrete values, continuous values, and probability values. The diversity of voting values ​​enables the system to adapt to decision-making needs at different granularities.

[0050] In this embodiment, specifically, while outputting the final decision result, the system attaches the final decision weight, self-continuity score, voting value, and deviation level of each agent to form a visualized decision tracing report.

[0051] The decision origination report enhances the interpretability of the system. The report presents the health status of each agent and its contribution to the final decision in the form of tables or charts, helping administrators understand the decision-making logic, quickly locate abnormal agents, and facilitate auditing and review.

[0052] Example 2

[0053] This embodiment uses a multi-agent collaborative investment decision-making system as an example to detail the complete process of a multi-agent dynamic weighted decision-making method based on self-continuity. The system includes three different types of agents: a risk-sensitive agent (Agent A), a value-investing agent (Agent B), and a trend-following agent (Agent C). Each agent has a built-in self-representation encoder. (Fine-tuned using a pre-trained BERT model) is used to map its own state into a 128-dimensional vector representation.

[0054] Step S1: Upon system startup, record the initial state of each agent. The initial state includes the agent's initial parameters, value anchors (such as risk aversion coefficient and value assessment model parameters), and behavioral patterns (such as historical trading strategies). This is achieved through a self-representation encoder. Calculate the initial state vector for each agent: Agent A: =[0.12,-0.34,…] (128-dimensional vector) Agent B: =[0.56,0.78,…] Agent C: =[-0.23,0.45,…] These initial vectors are stored in persistent storage as a benchmark for subsequent calculation of self-continuity scores.

[0055] Step S2: The system collects the current state of each agent every minute. The current state includes the agent's latest parameters, recent decision records, and market environment perception. This is also done through an encoder. Calculate the current state vector: At a certain decision time t, Agent A's current state vector is... =[0.11,-0.30,…], the current state vector of Agent B. =[0.60,0.80,…], the current state vector of Agent C =[-0.25,0.50,…].

[0056] Step S3: According to the formula The ego continuity score of each agent is calculated. In this embodiment, the squared Euclidean distances are calculated as follows: =0.98、 =0.85、 =0.92.

[0057] These scores indicate that Agent B deviates relatively significantly, but is still within the normal range.

[0058] Step S4: For the current decision topic "Whether to buy stock X", each agent outputs a voting value according to its own strategy. This embodiment uses discrete voting values: +1 indicates buying, -1 indicates selling, and 0 indicates waiting.

[0059] Agent A (risk-sensitive): The current market is highly volatile, so the output VA = -1. Agent B (Value Investor): Believes that stock X is undervalued, outputs VB=+1; Agent C (Trend Tracker): Detects an upward trend and outputs VC=+1.

[0060] Step S5: The system presets the base weights for each agent: Agent A is 0.8, Agent B is 1.2, and Agent C is 1.0 (base weights are set according to the importance of the role). According to the formula... Calculate the final decision weights: Agent A: =0.8 × 0.98 = 0.784; Agent B: =1.2 × 0.85 = 1.02; Agent C: =1.0 × 0.92 = 0.92; Step S6: Calculate a weighted average of all votes:

[0061] The system has preset decision thresholds: a buy threshold of 0.3 and a sell threshold of -0.3. Since the final score of 0.424 > 0.3, the system outputs a buy order.

[0062] Before making an actual decision, the system can activate a state pre-check mechanism. For example, if Agent B is ready to output a vote value of +1, the system, based on its current state and the decision to be output (buy), pre-analyzes the possible states after the decision and calculates an expected self-continuity score of 0.82 (lower than the current value of 0.85 but still higher than the first preset threshold of 0.7), thus allowing the output. If the expected score is lower than 0.7, the system will reject the decision and trigger a self-check process.

[0063] During the decision-making process, the system monitors the deviation degree Di in real time. Assume Agent B's deviation degree DB = 0.15, which is within the normal range. If Agent B's deviation degree rises to 0.35 (reaching the yellow threshold of 0.3) during subsequent operation, the system will log and reduce its base weight to 0.8; if it rises to 0.55 (reaching the red threshold of 0.5), the system will reject its vote and trigger a self-check; if it unfortunately rises to 0.85 (reaching the black threshold of 0.8), the system will forcibly roll back Agent B's state to its initial state and suspend its decision-making power.

[0064] If Agent B triggers a red-line interception, the system isolates it in a sandbox environment. In the sandbox, Agent B replays its most recent 10 decision paths and discovers that the main cause of the deviation was excessive updating of its valuation model parameters after a period of abnormal market volatility. The system compares the current decision path with the initial decision path, generates a discrepancy report, and attempts to revert the model parameters to the checkpoint before the deviation. After the correction, Agent B's deviation drops to 0.28, and the system restores its normal decision-making authority.

[0065] Self-representation encoder A neural network model is used, and its training process incorporates an ego continuity loss function.

[0066] In this embodiment, the base weights are dynamically adjusted according to the role: risk-sensitive agents can have their base weights increased appropriately in volatile markets, while value-investing agents can have their weights increased in undervalued markets. Voting values ​​can also be extended to continuous values; for example, Agent A outputting -0.8 indicates a strong sell, while Agent B outputting +0.6 indicates a moderate buy, making decision-making more nuanced.

[0067] After the final decision is generated, the system automatically produces a source tracing report, which includes information such as the final weight, self-continuity score, voting value, and deviation level of each agent, displayed in chart form. Administrators can use this report to quickly locate agents with potential risks and adjust system parameters.

[0068] Example 3

[0069] An electronic device includes at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform a self-continuity-based multi-agent dynamic weighted decision-making method.

[0070] Example 4

[0071] A computer-readable storage medium storing computer instructions for causing a computer to execute a multi-agent dynamic weighted decision-making method based on self-continuity.

[0072] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in the present invention, and these should all be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-agent dynamic weighted decision-making method based on self-continuity, characterized in that, Includes the following steps: Step S1: During the system initialization phase, collect and store the initial state vector of each agent. The initial state vector is generated by a self-representation encoder. It is obtained by encoding the initial parameters, value anchors, or behavioral patterns of the intelligent agent; Step S2: During system operation, periodically collect the current state of each agent. And through the self-representation encoder Map it to the current state vector ; Step S3: Calculate the ego continuity score for each agent. The self-continuity score is used to quantify the degree of deviation of the agent's current state from its initial state, and is calculated using the following formula: in, The square of the Euclidean distance. The value range is [0, 1]; Step S4: Obtain the voting value Vi for each agent regarding the current decision topic. The voting value Vi can be a discrete or continuous value, used to express the agent's preference for a certain action plan. Step S5: Based on the base weights of each agent Based on the self-continuity score continuityi, calculate the final decision weight for each agent. The calculation formula is: in, Determined by the pre-defined role of the agent; Step S6: Calculate the weighted average of all agent votes to obtain the final decision score F of the system. The calculation formula is as follows: Where n is the number of agents involved in the decision-making process; Step S7: Compare the final decision score F with the preset decision threshold, and output the final decision instruction.

2. The multi-agent dynamic weighted decision-making method based on self-continuity as described in claim 1, characterized in that, Before calculating the self-continuity score in step S3, a state pre-check step is also included: Before an agent is about to make a decision, the system uses the agent's current state as a basis. Given the decision to be output, simulate the possible future states that may result from the execution of the decision, and calculate the expected self-continuity score. If the self-continuity score is lower than the first preset threshold, the agent's current decision output is rejected, and a self-check process is triggered, requiring the agent to regenerate the decision. If the self-continuity score is lower than the second preset threshold and the second preset threshold is lower than the first preset threshold, then the agent's state is forcibly rolled back to the initial state, and its decision-making power is suspended until recalibration is completed.

3. The multi-agent dynamic weighted decision-making method based on self-continuity as described in claim 1, characterized in that, This includes a three-tiered safety line fuse mechanism, which is based on deviation. Trigger different levels of security responses; When deviation i ≥ Yellow line threshold and deviation i When the threshold value is less than the red line threshold, the system logs the event, reduces the final decision weight of the agent in subsequent decisions, and alerts the system administrator. The yellow line threshold value ranges from [0.2, 0.4]. When deviation i ≥ Red line threshold and deviation i When the deviation falls below the black line threshold, the system rejects the agent's decision, triggers a self-check process, requires the agent to regenerate its opinion, and temporarily freezes the agent's voting rights until its deviation drops below the red line threshold, which ranges from [0.4, 0.6]. When deviation i When the value is greater than or equal to the black line threshold, the system will force the agent's state to roll back to the initial state, suspend its speaking rights until stability is restored, and trigger the full system audit process. The value range of the black line threshold is [0.7, 0.9].

4. The multi-agent dynamic weighted decision-making method based on self-continuity as described in claim 3, characterized in that, The self-test process includes: The agent enters an isolated sandbox environment, replays its recent decision-making path in the sandbox, and analyzes the key decision nodes that lead to increased deviation. The system generates a difference report by comparing the current decision path with the initial decision path, and attempts to repair the agent by fine-tuning parameters or reverting to the checkpoint before the deviation. If the self-check process successfully reduces the deviation to below the red line threshold, the agent's normal decision-making power is restored; otherwise, the frozen state is maintained and manual intervention is awaited.

5. The multi-agent dynamic weighted decision-making method based on self-continuity as described in claim 1, characterized in that, The self-representation encoder For a trainable neural network model, an ego continuity loss function is introduced during its training process. The self-continuity loss function is defined as follows: During encoder training, the total loss function is: in, For the task loss function, This is a balance coefficient used to control the strength of the self-continuity constraint; it is achieved by minimizing... This allows the encoder to maintain sensitivity to changes in the agent's state while completing a specific task, reflecting the degree of drift in its intrinsic value.

6. The multi-agent dynamic weighted decision-making method based on self-continuity as described in claim 1, characterized in that, The basic weights are determined by the agent's role type, which includes, but is not limited to, risk-sensitive agents, value-investing agents, trend-following agents, and arbitrage agents.

7. The multi-agent dynamic weighted decision-making method based on self-continuity as described in claim 1, characterized in that, The voting values ​​can be expressed in various forms, including but not limited to binary discrete values, multivariate discrete values, continuous values, and probability values.

8. The multi-agent dynamic weighted decision-making method based on self-continuity as described in claim 1, characterized in that, Along with the final decision result, the system also includes the final decision weight, self-continuity score, voting value, and deviation level for each agent, forming a visualized decision tracing report.

9. An electronic device, characterized in that, The electronic device includes at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the self-continuity-based multi-agent dynamic weighted decision-making method according to any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the self-continuity-based multi-agent dynamic weighted decision-making method as described in any one of claims 1-8.

Images