A multi-agent view evolution simulation system and method based on cognitive bias

Abstract

The application discloses a kind of based on cognitive bias multi-agent view evolution simulation system and method, the method includes: constructing directed weighted influence graph to characterize the asymmetric influence relationship between agents in social network, and each agent is allocated initial view state;Each agent is configured with personalized cognitive bias function, the cognitive bias function dynamically adjusts influence weight according to the view difference between influencer and the influenced person;Based on the directed weighted influence graph and the cognitive bias function, nonlinear view updating rule is executed, and the agent view state of next time step is generated;The present application aims to solve the technical problems that the simulation results are disconnected with real social dynamics due to the lack of individual cognitive heterogeneity, the simplification of nonlinear influence mechanism, the lack of unified framework for cognitive bias modeling, the lack of convergence criterion and the insufficient system stability in the existing social network view evolution model.

Description

Technical Field

[0001] This invention relates to the technical field of information technology, specifically to a multi-agent perspective evolution simulation system and method based on cognitive bias. Background Technology

[0002] In multi-agent social network opinion evolution simulation, the core technical challenge lies in accurately characterizing the nonlinear impact of opinion divergence on opinion updates while preserving individual cognitive heterogeneity, and on this basis, predicting whether the group tends towards consensus or polarization. This problem involves modeling social relations as a directed weighted influence graph, but existing models such as the DeGroot framework usually assume that all individuals follow the same linear weighted average rule, which leads to an oversimplification of the opinion update process and makes it difficult to fully reflect the complex dynamic behaviors caused by differences in psychological mechanisms in real society.

[0003] Furthermore, while some improvement schemes attempt to reflect specific biases such as conformity and obedience to authority through threshold filtering or fixed weight adjustment when introducing cognitive biases, they lack a unified mathematical representation framework, making it difficult for the model to adapt to diverse cognitive patterns such as confirmation bias and backfire effect. Moreover, they usually assume that the same bias works in the same way for all individuals, ignoring the differences in cognitive responses that individuals may show to different influencers.

[0004] Furthermore, existing methods mostly rely on full time-step iterations to observe the final state, without establishing prior convergence criteria based on network topology and deviation function characteristics. This makes it difficult to efficiently predict whether the system has reached a consensus when facing large-scale or non-strongly connected networks, thus limiting its decision support capabilities in public opinion intervention or policy inference.

[0005] Ultimately, when dealing with nonlinear update rules, due to the lack of rigorous mathematical analysis tools, some models may exhibit viewpoint oscillations or long-term polarization under specific parameter configurations. Furthermore, existing technologies do not provide clear constraints on the continuity and action boundaries of the deviation function, making it difficult to guarantee system stability and affecting the reliability and interpretability of simulation results.

[0006] This comprehensive problem spans the entire process from network modeling and bias embedding to dynamic evolution and convergence determination, involving the cross-integration of graph theory, nonlinear dynamics, and social psychology, and directly affecting the real-world relevance and application value of viewpoint evolution simulation.

[0007] More critically, existing technologies, after introducing cognitive biases that lead to system nonlinearity, lack effective theoretical analysis tools (such as Markov chain failures), making it impossible to rigorously prove the system's convergence. Furthermore, existing models heavily rely on the unrealistic assumption that the network is a strongly connected graph, failing to handle the complex topologies prevalent in real-world societies, which contain information silos or unidirectional influence chains. This results in severely inadequate consensus prediction capabilities in non-strongly connected networks. These shortcomings collectively limit the practical application value of traditional models in complex social public opinion simulation and intervention decision-making. Summary of the Invention

[0008] This invention provides a multi-agent viewpoint evolution simulation system and method based on cognitive bias, aiming to solve the technical problems in existing social network viewpoint evolution models, such as the lack of individual cognitive heterogeneity, simplification of nonlinear influence mechanisms, lack of a unified framework for cognitive bias modeling, lack of convergence criteria, and insufficient system stability, which lead to the disconnect between simulation results and real social dynamics.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0010] A multi-agent viewpoint evolution simulation method based on cognitive bias includes: constructing a directed weighted influence graph to characterize the asymmetric influence relationship between agents in a social network, and assigning an initial viewpoint state to each agent; configuring a personalized cognitive bias function for each agent, wherein the cognitive bias function dynamically adjusts the influence weights according to the viewpoint differences between the influencing agent and the influenced agent; executing a nonlinear viewpoint update rule based on the directed weighted influence graph and the cognitive bias function to generate the agent viewpoint state for the next time step; after each viewpoint update, calculating the dispersion index of the group viewpoint distribution, and combining the network topology characteristics and the continuity boundary conditions of the cognitive bias function to determine whether the system meets a preset convergence criterion; if the convergence criterion is met, terminating the simulation and outputting the final state of the group evolution; if the convergence criterion is not met and the number of iterations has not reached a preset upper limit, returning to execute the nonlinear viewpoint update step; if the number of iterations reaches the preset upper limit and the convergence criterion is still not met, marking the system as being in an oscillation or polarization state and outputting the corresponding evolution trajectory.

[0011] In one aspect of this disclosure, the step of constructing a directed weighted influence graph to characterize the asymmetric influence relationships among agents in a social network, and assigning an initial viewpoint state to each agent, includes:

[0012] The system receives an input social relationship dataset, which contains agent identities and the strength of their mutual influence.

[0013] Based on the social relationship dataset, an adjacency matrix is ​​constructed, where the matrix elements represent the influence weights from one agent to another.

[0014] The adjacency matrix is ​​normalized so that the sum of the outgoing edge weights of each agent is equal to a unit value, forming a row random matrix;

[0015] Each agent is assigned an initial viewpoint value, which is randomly sampled from a preset viewpoint space interval.

[0016] The row random matrix is ​​combined with the initial viewpoint value to form a joint representation of the initial state vector and the influence graph structure;

[0017] Verify whether the influence graph satisfies the weak connectivity condition. If it does not, record the subgraph partitioning information for subsequent grouping analysis.

[0018] The influence graph structure and initial viewpoint state are stored in the simulation environment memory as the starting point for evolution.

[0019] In one aspect of this disclosure, the step of configuring a personalized cognitive bias function for each agent, wherein the cognitive bias function dynamically adjusts the influence weights based on the differences in viewpoints between the influencer and the influenced, includes:

[0020] Define a deviation type identifier for each agent, where the deviation type includes one or more combinations of confirmation bias, backfire effect, conformity tendency, or authority obedience.

[0021] Based on the deviation type identifier, select the corresponding mathematical expression as the basic function from the preset deviation function library;

[0022] Configure an individual-specific set of parameters for the base function, the set of parameters being obtained through user input or learned from historical behavioral data;

[0023] The individual-specific parameter set is embedded into the basic function to generate a personalized cognitive bias function for the agent.

[0024] During the simulation, when the agent receives opinion signals from the influencer, it calculates the difference between the two opinions;

[0025] Substitute the aforementioned opinion difference into the personalized cognitive bias function to output the corrected dynamic influence weight;

[0026] The dynamic influence weights are multiplied by the static weights in the original influence diagram to generate effective influence weights for viewpoint updates.

[0027] In one aspect of this disclosure, the step of executing a nonlinear view update rule based on the directed weighted influence graph and the cognitive bias function to generate the agent's view state for the next time step includes:

[0028] Iterate through all agents and calculate their viewpoint values ​​for the next time step in turn;

[0029] For the current agent, collect the view values ​​of all its neighboring agents at the current time step;

[0030] Based on the personalized cognitive bias function, calculate the acceptance weight of the current agent for each neighbor's viewpoint.

[0031] The acceptance weights are multiplied element-wise with the original connection weights in the influence graph to generate a comprehensive influence weight vector.

[0032] The comprehensive influence weight vector is normalized to ensure that the sum of the weights is a unit value;

[0033] The normalized weight vector is summed with the neighbor view values ​​to obtain the linear prediction value;

[0034] The linear prediction value is input into a preset nonlinear activation function to generate the final viewpoint value for the next time step.

[0035] Synchronize and update the new viewpoints of all agents to the simulation state cache to complete one full iteration.

[0036] In one aspect of this disclosure, the step of calculating the dispersion index of the group's viewpoint distribution after each viewpoint update, and determining whether the system satisfies a preset convergence criterion by combining the network topology characteristics and the continuity boundary conditions of the cognitive bias function, includes:

[0037] Calculate the standard deviation or entropy of all agent viewpoints at the current time step as an indicator of population dispersion;

[0038] Extract topological features such as algebraic connectivity, size of the largest strongly connected component, and average path length of the current influence graph;

[0039] Verify whether all personalized cognitive bias functions satisfy the Lipschitz continuity condition within their domain;

[0040] The discreteness index, topological features, and continuity verification results are input into a preset convergence criterion logic unit;

[0041] The convergence criterion logic unit determines whether the dispersion index is lower than a preset threshold, and the topology supports information propagation, and the deviation function satisfies stability constraints.

[0042] If all three conditions are met, the system is determined to have converged to a consensus or stable polarization state.

[0043] If any condition is not met, the system is determined not to have converged, and the next iteration continues.

[0044] In one aspect of this disclosure, the step of terminating the simulation and outputting the final state of population evolution if the convergence criterion is satisfied includes:

[0045] Stop the time step iteration loop and freeze the current view state of all agents;

[0046] The number of clusters and the location of the center of each cluster in the distribution of group opinions;

[0047] Identify whether there are multiple significantly separated clusters of viewpoints to determine whether the system is in a state of consensus, mild divergence, or extreme polarization;

[0048] The final state, evolution path, and key criteria parameters are packaged into a simulation report;

[0049] The visualization engine renders the evolution of viewpoints into dynamic heatmaps or network animations.

[0050] Write the simulation report and visualization data to persistent storage media for later analysis;

[0051] The external interface is triggered to notify the user that the simulation task has been completed.

[0052] In another aspect, this disclosure also relates to a multi-agent perspective evolution simulation system based on cognitive bias, comprising:

[0053] The network construction module is configured to construct a directed weighted influence graph to characterize the asymmetric influence relationships between agents in a social network and to assign an initial viewpoint state to each agent.

[0054] The bias configuration module is configured to configure a personalized cognitive bias function for each agent, wherein the cognitive bias function dynamically adjusts the influence weight based on the difference in viewpoints between the influencer and the influenced;

[0055] The viewpoint update module is configured to perform the step of executing a nonlinear viewpoint update rule based on the directed weighted influence graph and the cognitive bias function to generate the agent's viewpoint state for the next time step;

[0056] The convergence discrimination module is configured to calculate the dispersion index of the group's viewpoint distribution after each viewpoint update, and determine whether the system meets the preset convergence criterion by combining the network topology characteristics and the continuity boundary conditions of the cognitive bias function.

[0057] The state output module is configured to terminate the simulation and output the final state of the population evolution if the convergence criterion is met.

[0058] The oscillation monitoring module is configured to record the system oscillation period or polarization mode and output an abnormal evolution log when the number of iterations reaches a preset upper limit and still fails to converge.

[0059] The parameter management module is configured to uniformly maintain the configuration parameters of the cognitive bias function library, topological feature extraction algorithm, and convergence criterion logic unit.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] This invention introduces a personalized cognitive bias function to achieve refined modeling of individual heterogeneous responses under different psychological mechanisms, overcoming the oversimplification of complex social cognitive processes by traditional linear models. By combining nonlinear update rules with dynamic weight adjustments, it accurately characterizes the non-monotonic impact of opinion divergence on opinion adoption, and can reproduce real social phenomena such as confirmation bias and backfire. By establishing a prior convergence criterion that integrates network topology, bias function continuity, and group dispersion, it avoids the computational overhead of full-cycle simulation of large-scale networks, significantly improving simulation efficiency and decision support capabilities. By imposing strict mathematical constraints on the bias function, it ensures the dynamic stability of the system in long-term evolution, preventing meaningless opinion oscillations, thereby improving the reliability and interpretability of simulation results. By introducing strict mathematical constraints (such as continuity) and applying real analysis principles, it theoretically guarantees the stability and convergence of nonlinear systems. The overall technical solution connects the complete process from network construction, bias embedding, dynamic evolution to state discrimination, providing a high-fidelity and verifiable computational experimental platform for public opinion analysis, policy projection, and social intervention. Attached Figure Description

[0062] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0063] Figure 1 This is a schematic diagram of the four characteristic regions (M, R, B, I) of the cognitive bias function in an embodiment of the present invention.

[0064] Figure 2 This is a schematic diagram of the curves of five typical cognitive bias functions in the opinion difference interval [-1, 1] in the embodiments of the present invention.

[0065] Figure 3 This is a schematic diagram of the directed weighted influence graph structure in the simulation example of vaccine safety disputes applied to this invention. Detailed Implementation

[0066] The present invention will be further described below with reference to embodiments. These embodiments are merely some, not all, of the embodiments of the present invention. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the protection scope of the present invention.

[0067] Example 1

[0068] Please see Figures 1-3 As shown, this embodiment discloses a multi-agent viewpoint evolution simulation method based on cognitive bias, including: constructing a directed weighted influence graph to characterize the asymmetric influence relationship between agents in a social network, and assigning an initial viewpoint state to each agent; configuring a personalized cognitive bias function for each agent, wherein the cognitive bias function dynamically adjusts the influence weight according to the viewpoint difference between the influencer and the influenced; executing a nonlinear viewpoint update rule based on the directed weighted influence graph and the cognitive bias function to generate the agent viewpoint state for the next time step; after each viewpoint update, calculating the dispersion index of the group viewpoint distribution, and combining the network topology characteristics and the continuity boundary conditions of the cognitive bias function to determine whether the system meets the preset convergence criterion; if the convergence criterion is met, terminating the simulation and outputting the final state of the group evolution; if the convergence criterion is not met and the number of iterations has not reached the preset upper limit, returning to execute the nonlinear viewpoint update step; if the number of iterations reaches the preset upper limit and the convergence criterion is still not met, marking the system as being in an oscillation or polarization state and outputting the corresponding evolution trajectory.

[0069] This invention constructs a directed weighted influence graph to characterize the asymmetric social influence relationships between individuals and assigns a personalized cognitive bias function to each agent, enabling it to dynamically adjust its acceptance of different neighbors based on differing viewpoints. Building upon this, a nonlinear update mechanism integrating static network weights and dynamic bias correction is employed to model complex psychological phenomena such as confirmation bias and backfire. Simultaneously, after each iteration, a priori criteria incorporating group dispersion, network connectivity, and the Lipschitz constant are introduced to pre-determine system convergence and avoid ineffective iterations. Finally, structured output of the evolutionary final state or anomalous trajectories provides interpretable and verifiable simulation results for public opinion analysis. This invention significantly improves the ability of viewpoint evolution models to characterize the heterogeneity of real-world social cognition, enhances the accuracy of reproducing nonlinear dynamics, and improves computational efficiency in large-scale networks.

[0070] Example 2

[0071] Please see Figures 1-3 As shown, this embodiment is a further optimization based on Embodiment 1. In this embodiment, the multi-agent viewpoint evolution simulation method based on cognitive bias may specifically include:

[0072] S101. Construct a directed weighted influence graph to characterize the asymmetric influence relationship between agents in a social network, and assign an initial viewpoint state to each agent.

[0073] The system receives an input social relationship dataset, which includes agent identities and their mutual influence strengths; based on the social relationship dataset, it constructs an adjacency matrix. , of which elements Indicates from the intelligent agent Pointing to intelligent agents The influence weights; the adjacency matrix is ​​row-normalized so that... For all Established, forming a row random matrix; for each agent Assign initial viewpoint values This value is generated by random sampling from a uniform distribution; the normalized adjacency matrix is ​​then compared with the initial view vector. Combine the elements to form a joint state representation; verify whether the influence graph satisfies weak connectivity, and if not, record its strongly connected component partitioning; store the graph structure and initial state in the simulation environment memory as the starting point for evolution.

[0074] For example, suppose the input social relationship data contains 5 agents, and the original form of their influence strength matrix is ​​shown in the table below:

[0075] After normalizing each row, a row random matrix is ​​obtained. For example, the sum of the outgoing edge weights of agent B is... No adjustment is needed; however, if the sum of a row is not 1, it will be scaled proportionally. The initial viewpoint value is set to... The system detected that the graph contains a maximally strongly connected component. C and D form the downstream subgraph. The above data structure is serialized and stored in a memory buffer for subsequent modules to access.

[0076] S102. Configure a personalized cognitive bias function for each agent, wherein the cognitive bias function dynamically adjusts the influence weight according to the difference in viewpoints between the influencer and the influenced.

[0077] like Figure 1 As shown, in this step, the cognitive bias function is... Defined as differences in viewpoints The mappings ∈ [-1.1] are divided into the following four regions based on their teaching characteristics:

[0078] Compliant region M: · ≥0 and This indicates that the agent blindly accepts or amplifies external viewpoints. Within this region, the agent will accept and may amplify the influence of external viewpoints, simulating blind obedience to authority or blindly following the crowd.

[0079] Acceptance-resistance region (R): · ≥0 and This indicates that the agent conditionally accepts external viewpoints, and the degree of acceptance changes dynamically with the differences in viewpoints. Within this region, the agent conditionally accepts external viewpoints, and the degree of acceptance is non-linearly related to the differences in viewpoints, simulating common psychological phenomena such as confirmation bias.

[0080] Backfire type area (B): · This indicates that the agent rejects opposing viewpoints and reinforces its own original viewpoints. Within this region, the agent rejects opposing viewpoints and reinforces its own original viewpoints, simulating the backfire effect.

[0081] Isolated region (I): =0. This indicates that the agent completely ignores external viewpoints. Within this region, the agent completely ignores all external viewpoints and is in an information silo state.

[0082] The personalized cognitive bias function It is configured to belong to at least one of the above regions.

[0083] Based on this classification framework, this invention constructs an extensible deviation function library. The following are examples of several typical preset functions:

[0084] Unbiased function: =1; belongs to the M region, degenerates into a classical linear model.

[0085] Confirmation deviation function: , where k>0, belongs to region R.

[0086] Backfire effect function: It belongs to region B.

[0087] Authority bias function: ,in This belongs to region M and simulates the unconditional amplification of authoritative viewpoints.

[0088] Isolated deviation function: =0, belongs to region I.

[0089] For each agent, a bias type identifier is defined, including one or more combinations of confirmation bias, backfire effect, conformity tendency, or authority obedience. Based on the bias type identifier, a corresponding mathematical expression is selected as a base function from a pre-defined bias function library. An individual-specific parameter set is configured for the base function, obtained through user input or learned from historical behavioral data. The individual-specific parameter set is embedded into the base function to generate a personalized cognitive bias function for that agent. During simulation, when the agent... Receive from neighbors viewpoint signals At that time, calculate the difference in opinions. Substitute into the personalized cognitive bias function Output correction factor ;Will Compared with the static weights in the original influence diagram Multiply to generate effective influence weights .

[0090] For example, agent A is configured to have an acknowledgment bias, the bias function of which is defined as follows: ,in The individual-specific attenuation coefficient; agent B is configured for backfire effect, whose function is: ,in At time step Agent A receives opinions from B. My own opinion is Therefore Substitute have to Therefore, effective weights This process is performed in parallel on all neighbors, generating a complete dynamic weight matrix.

[0091] S103. Based on the directed weighted influence graph and the cognitive bias function, execute the nonlinear view update rule to generate the agent's view state for the next time step.

[0092] Iterate through all agents and calculate their viewpoint values ​​for the next time step; for the current agent $i$, collect all its incoming neighbors. Viewpoint at the current time step Based on the personalized cognitive bias function, calculate respectively ;Will With original connection weights Element-wise multiplication generates a comprehensive influence weight vector. ;right After normalization, we get ; Calculate linear predicted values ;Will Input a preset non-linear activation function Generate the final next time step view value. All new viewpoints are synchronously written to the state cache, completing one full iteration.

[0093] For example, a variant of the Sigmoid activation function is chosen. ,in Control the steepness. Assume agent A is in... Normalized predicted values ,but This nonlinear transformation ensures that the viewpoint value is always within the range specified in the original text. Within the range, the sensitivity of the middle region is amplified, and extreme value mutations are suppressed.

[0094] S104. After each opinion update, calculate the dispersion index of the group opinion distribution, and combine the network topology characteristics and the continuity boundary conditions of the cognitive bias function to determine whether the system meets the preset convergence criterion.

[0095] During the iteration process, the following topology analysis and prediction are performed periodically:

[0096] a) Perform strong connected component decomposition on the directed weighted influence graph to identify all source components (i.e., strong connected components without incoming edges).

[0097] b) Monitor the convergence of the viewpoints of the agents within each source component. If all source components have converged to their respective stable values, and these stable values ​​are the same, then it can be determined that the entire system will tend towards global consensus.

[0098] c) This prediction result can be used in conjunction with criteria based on the dispersion index to terminate the simulation in advance or to provide early warning of polarization state.

[0099] Calculate the standard deviation of all agent viewpoints at the current time step. As a measure of dispersion; extract the algebraic connectivity of the current influence graph. (i.e., the second smallest eigenvalue of the Laplacian matrix), the size of the largest strongly connected component and average shortest path length ; Validate all personalized cognitive bias functions In its domain Does the above satisfy Lipschitz continuity, i.e., does a constant exist? For any Establish; will as well as Input a preset convergence criterion logic unit; the logic unit determines whether the following conditions are met simultaneously: , like ,and (like If all three conditions are met, the system is considered to have converged; otherwise, the iteration continues.

[0100] For example, in hour, Furthermore, since the Lipschitz constants of all deviation functions are less than 1.4, the convergence criterion is satisfied, and the system terminates.

[0101] S105. If the convergence criterion is met, the simulation is terminated and the final state of population evolution is output.

[0102] The time-step iteration loop is stopped, freezing the current viewpoint states of all agents; the K-means clustering algorithm is used to group the final viewpoint values, and the number of clusters is set. Alternatively, 3. Count the number of clusters and the location of each cluster center; if the largest cluster accounts for more than 90%, it is considered a consensus state; if there are two clusters with a center-to-center distance greater than 0.6, it is considered extreme polarization; otherwise, it is considered mild divergence; the final state, evolution path, and key criterion parameters (such as...) are then analyzed. Package and generate a JSON-formatted simulation report; render the evolution trajectory of opinions as a dynamic heatmap using a visualization engine, with the horizontal axis representing time steps, the vertical axis representing agent IDs, and color depth indicating opinion values; write the report and visualization data to persistent storage; trigger a RESTful API to notify the user that the task is complete.

[0103] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principles of this invention are further supplemented below with a specific application scenario.

[0104] When deploying the method of this invention in a public opinion intervention decision support system, the network construction module first reads user interaction logs from real social media platforms. These logs contain user IDs and the implied influence strength of behaviors such as forwarding, commenting, and liking. Based on this, the system generates a directed weighted influence graph, where nodes correspond to agents and edge weights... user right The interaction frequency and emotional polarity are weighted and calculated, and then row-normalized to form a row-random matrix, such as... Figure 2 The structure is shown on the left side; initial viewpoint value. Then, the sentiment score based on the user's historical posts is mapped to... The interval is used to complete the initialization of step S101.

[0105] Then, in stage S102, the bias configuration module assigns a cognitive bias type to each agent based on user profile data: for example, a user with a conservative political stance is labeled as having "confirmation bias," and their personalized function... attenuation coefficient in The output is derived from a learning model of their past information filtering behavior; while users easily swayed by extreme rhetoric are assigned a "backfire effect" function. ,in This indicates the increased degree of rejection of heterogeneous viewpoints; such as... Figure 2 As shown, when agent A (confirmation bias) receives information from neighbor B with significantly different viewpoints, its effective weight... It is significantly compressed, but if information is received from E with similar viewpoints, then The weights have almost no decay, thus enabling dynamic modeling of the "echo chamber effect".

[0106] When performing nonlinear updates in S103, the view update module updates each agent... Collect current opinions from its neighbors. and using the calculated Construct a temporary weight vector; this vector, after normalization, is used to obtain a weighted average. Then input the Sigmoid activation function. ; here The setting makes when When the output value is close to 0 or 1, the change is gradual to prevent sudden shifts in opinion. However, in the intermediate region (e.g., 0.3–0.7), minute differences are amplified to simulate the rapid shifts in opinion near the "critical point" in real-world society. Figure 2 As shown by the curve on the right, this nonlinear mechanism is key to reproducing viewpoint polarization or consensus shifts.

[0107] Entering the S104 convergence determination stage, the convergence determination module calls the graph theory analysis subroutine to calculate the second smallest eigenvalue of the Laplacian matrix after each iteration. ,like This demonstrates that the network possesses global information dissemination capabilities; it also calculates the standard deviation of group opinions. A value below 0.02 indicates a high degree of concentration of opinions; furthermore, the system initializes all... Perform Lipschitz constant Upper bound estimate (e.g. for) Its derivative has a maximum absolute value of 2, therefore However, in this example, due to the use of a smooth exponential function and controlled parameters, the actual... ;like Figure 3 As shown, when three conditions When both conditions are met, the logic unit determines that the system has converged and there is no need to continue iterating, thereby avoiding redundant calculations on large networks.

[0108] Finally, in the S105 state output phase, the system freezes the current view vector and calls a clustering algorithm to identify view clusters: if 95% of the agents cluster in... If the interval is consistent, it is considered a consensus; if two clusters are located in different intervals, it is considered a consensus. and Furthermore, if the center distance is greater than 0.6, it is marked as extreme polarization; the final generated JSON report includes the final state view of each agent, evolution path snapshot, convergence time step, and criterion parameter values, and is rendered by a visualization engine. Figure 1 The heatmap shown has time steps (0–50) on the horizontal axis and agent IDs (A–E) on the vertical axis. The color gradually changes from blue (0.0) to red (1.0), which intuitively shows that A, B, and E quickly converge within strongly connected components, while C and D lag or deviate due to being downstream and affected by high bias. This provides policymakers with a basis for identifying intervention windows.

[0109] In some optional embodiments, the specific steps in the application scenario of extrapolating the social acceptance of epidemic control policies are as follows:

[0110] S1: Scene initialization and network construction.

[0111] Specifically, in this scenario, the system simulates the evolution of acceptance of a "mandatory home quarantine" policy in a community of 100 agents. The network construction module generates a weakly connected directed weighted influence graph based on real social data, where nodes represent individuals and edge weights combine interaction frequency, relationship intimacy, and professional authority. The initial opinion state is set based on questionnaire data, mapping the degree of acceptance of the policy (0 for strong opposition, 1 for complete support), presenting an initial distribution of cautious supporters, strong opponents, and the middle group, laying a complex and realistic starting point for the evolution simulation.

[0112] S2: Configure personalized cognitive bias functions.

[0113] Specifically, the bias configuration module assigns different psychological cognitive modes to individuals based on their social attributes and historical behavior. Medical experts are assigned an authority obedience bias, reinforcing their influence within their professional circles; individuals impacted by the pandemic are assigned a backfire effect function, rejecting views supporting control and reinforcing their own stance; those with limited information channels are assigned a confirmation bias function, tending to accept similar viewpoints and filter contradictory information, simulating an "information cocoon"; while ordinary residents, easily influenced, exhibit a herd mentality. This refined bias configuration is the core of reproducing the heterogeneity of real-world social cognition.

[0114] S3: Perform a nonlinear viewpoint evolution simulation.

[0115] Specifically, the viewpoint update module initiates iterations. In the initial simulation phase, the internal "echo chambers" in the network quickly converge due to confirmation bias, but barriers rise between chambers due to viewpoint differences and the backfire effect, resulting in multipolarity. The "centrists" at pivotal positions are pulled in different directions by multiple extreme viewpoints. The backfire effect may cause some individual viewpoints to move to extremes, widening the divergence; while the nonlinear activation function ensures that viewpoint values ​​change smoothly in the [0,1] interval and amplifies sensitivity in the intermediate region, simulating the rapid shift of viewpoint "critical points".

[0116] S4: Convergence judgment and process monitoring.

[0117] Specifically, the convergence discrimination module calculates indicators such as the dispersion of group viewpoints and the algebraic connectivity of the network after each iteration, and verifies the Lipschitz continuity of all cognitive bias functions. The simulation may lead to two typical outcomes: if the network connectivity is good and the extreme bias is not strong, the system may converge to a moderate consensus; if the opposing community is close-knit and the backfire effect is strong, a stable polarized state may be formed, and the system is marked as polarized after reaching the iteration limit, and detailed information of the opposing viewpoint clusters is recorded.

[0118] S5: Results output and decision support.

[0119] Specifically, the status output module generates structured reports and visual charts. If consensus is reached, the report will identify key influencers and time windows; if polarization occurs, it will clearly identify opposing groups and key nodes. More importantly, the system can serve as a "policy testing ground," allowing simulators to modify parameters (such as weakening the backfire effect of specific groups or increasing bridging connections) and conduct intervention tests. This allows for a quantitative evaluation of the effectiveness of different communication strategies in breaking polarization and promoting consensus, thereby providing precise and verifiable decision support for public opinion guidance and policy communication in real-world situations.

[0120] All content not described in detail in this specification belongs to existing technology known to those skilled in the art, and the implementation of each algorithm module can be completed based on common open-source libraries such as Python NetworkX, NumPy, and Scikit-learn. This technical solution does not specifically limit the hardware platform and can be deployed and run in a general server or cloud computing environment. The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0121] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-agent viewpoint evolution simulation method based on cognitive bias, characterized in that, include: Construct a directed weighted influence graph to characterize the asymmetric influence relationships between agents in a social network, and assign an initial viewpoint state to each agent; Each agent is configured with a personalized cognitive bias function, which dynamically adjusts the influence weights based on the differences in viewpoints between the influencer and the influenced. Based on the directed weighted influence graph and the cognitive bias function, a nonlinear view update rule is executed to generate the agent's view state for the next time step; After each opinion update, the dispersion index of the group opinion distribution is calculated, and combined with the network topology characteristics and the continuity boundary conditions of the cognitive bias function, it is determined whether the system meets the preset convergence criterion. If the convergence criterion is met, the simulation is terminated and the final state of population evolution is output. If the convergence criterion is not met and the number of iterations has not reached the preset upper limit, then return to execute the nonlinear view update step; If the convergence criterion is not met even after the number of iterations reaches the preset upper limit, the system is marked as being in an oscillation or polarization state and the corresponding evolution trajectory is output.

2. The multi-agent viewpoint evolution simulation method based on cognitive bias according to claim 1, characterized in that, Construct a directed weighted influence graph to characterize the asymmetric influence relationships between agents in a social network, and assign an initial viewpoint state to each agent; Each agent is configured with a personalized cognitive bias function, which dynamically adjusts the influence weights based on the differences in viewpoints between the influencer and the influenced. Based on the directed weighted influence graph and the cognitive bias function, a nonlinear view update rule is executed to generate the agent's view state for the next time step; After each opinion update, the dispersion index of the group opinion distribution is calculated, and combined with the network topology characteristics and the continuity boundary conditions of the cognitive bias function, it is determined whether the system meets the preset convergence criterion. If the convergence criterion is met, the simulation is terminated and the final state of population evolution is output. If the convergence criterion is not met and the number of iterations has not reached the preset upper limit, then return to execute the nonlinear view update step; If the convergence criterion is not met even after the number of iterations reaches the preset upper limit, the system is marked as being in an oscillation or polarization state and the corresponding evolution trajectory is output.

3. The multi-agent viewpoint evolution simulation method based on cognitive bias according to claim 1, characterized in that: The step of configuring a personalized cognitive bias function for each agent, wherein the cognitive bias function dynamically adjusts the influence weights based on the differences in viewpoints between the influencer and the influenced, includes: Define a deviation type identifier for each agent, where the deviation type includes one or more combinations of confirmation bias, backfire effect, conformity tendency, or authority obedience. Based on the deviation type identifier, select the corresponding mathematical expression as the basic function from the preset deviation function library; Configure an individual-specific set of parameters for the base function, the set of parameters being obtained through user input or learned from historical behavioral data; The individual-specific parameter set is embedded into the basic function to generate a personalized cognitive bias function for the agent. During the simulation, when the agent receives opinion signals from the influencer, it calculates the difference between the two opinions; Substitute the aforementioned opinion difference into the personalized cognitive bias function to output the corrected dynamic influence weight; The dynamic influence weights are multiplied by the static weights in the original influence diagram to generate effective influence weights for viewpoint updates.

4. The multi-agent viewpoint evolution simulation method based on cognitive bias according to claim 1, characterized in that: The step of generating the agent's viewpoint state for the next time step by executing a nonlinear viewpoint update rule based on the directed weighted influence graph and the cognitive bias function includes: Iterate through all agents and calculate their viewpoint values ​​for the next time step in turn; For the current agent, collect the viewpoints of all its in-neighbor agents at the current time step; Based on the personalized cognitive bias function, calculate the acceptance weight of the current agent for each neighbor's viewpoint. The acceptance weights are multiplied element-wise with the original connection weights in the influence graph to generate a comprehensive influence weight vector. The comprehensive influence weight vector is normalized to ensure that the sum of the weights is 1; The normalized weight vector is summed with the neighbor view values ​​to obtain the linear prediction value; The linear prediction value is input into a preset nonlinear activation function to generate the final viewpoint value for the next time step. Synchronize and update the new viewpoints of all agents to the simulation state cache to complete one full iteration.

5. The multi-agent viewpoint evolution simulation method based on cognitive bias according to claim 1, characterized in that: The step of calculating the dispersion index of the group's viewpoint distribution after each viewpoint update, and determining whether the system meets the preset convergence criterion by combining the network topology characteristics and the continuity boundary conditions of the cognitive bias function, includes: Calculate the standard deviation of all agent viewpoints at the current time step as an indicator of group dispersion. Extract the algebraic connectivity, the size of the largest strongly connected component, and the average shortest path length of the current influence graph as topological features; Verify that all personalized cognitive bias functions are within their domain. Does it satisfy the Lipschitz continuity condition? The discreteness index, topological features, and continuity verification results are input into a preset convergence criterion logic unit; The convergence criterion logic unit determines whether the dispersion index is lower than the first preset threshold, and whether the algebraic connectivity is greater than the second preset threshold, and whether the Lipschitz constants of all cognitive bias functions are less than the third preset threshold. If all three conditions are met, the system is determined to have converged to a consensus or stable polarization state. If any condition is not met, the system is determined not to have converged, and the next iteration continues.

6. The multi-agent viewpoint evolution simulation method based on cognitive bias according to claim 1, characterized in that: The step of terminating the simulation and outputting the final state of population evolution if the convergence criterion is met includes: Stop the time step iteration loop and freeze the current view state of all agents; The final opinion values ​​were grouped using the K-means clustering algorithm, with the number of clusters set to 2 or 3, and the number of clusters and the location of each cluster center were counted. If the largest cluster accounts for more than 90%, it is considered a consensus state; if there are two clusters and the distance between their centers is greater than 0.6, it is considered an extreme polarization state; otherwise, it is considered a mild divergence state. The final state, evolution path, and key criteria parameters are packaged into a JSON format simulation report; The visualization engine renders the evolution trajectory of opinions into a dynamic heatmap, with the horizontal axis representing time steps and the vertical axis representing agent identifiers, and the color intensity representing opinion values. Write the simulation report and visualization data to persistent storage media; The external interface is triggered to notify the user that the simulation task has been completed.

7. The multi-agent viewpoint evolution simulation method based on cognitive bias according to claim 1, characterized in that: The nonlinear activation function is: ,in This is the preset steepness control parameter, with a value of 4.

0.

8. The multi-agent viewpoint evolution simulation method based on cognitive bias according to claim 1, characterized in that: The cognitive bias function corresponding to the confirmation bias is: ,in The individual-specific attenuation coefficient is given; the cognitive bias function corresponding to the backfire effect is... ,in This is the individual-specific enhancement coefficient.

9. A multi-agent viewpoint evolution simulation system based on cognitive bias, characterized in that: The system includes: The network construction module is configured to perform the steps described in claim 1: constructing a directed weighted influence graph to characterize the asymmetric influence relationships between agents in a social network, and assigning an initial viewpoint state to each agent. The bias configuration module is configured to perform the step described in claim 1 of configuring a personalized cognitive bias function for each agent, wherein the cognitive bias function dynamically adjusts the influence weights based on the differences in viewpoints between the influencer and the influenced; The viewpoint update module is configured to execute the steps described in claim 1, which involve executing a nonlinear viewpoint update rule based on the directed weighted influence graph and the cognitive bias function, to generate the agent's viewpoint state for the next time step. The convergence discrimination module is configured to perform the steps described in claim 1, which are to calculate the dispersion index of the group's viewpoint distribution after each viewpoint update, and to determine whether the system meets the preset convergence criteria by combining the network topology characteristics and the continuity boundary conditions of the cognitive bias function. The state output module is configured to execute the step described in claim 1, which is to terminate the simulation and output the final state of the population evolution if the convergence criterion is met; The oscillation monitoring module is configured to record the system oscillation period or polarization mode and output an abnormal evolution log when the number of iterations reaches a preset upper limit and still fails to converge. The parameter management module is configured to uniformly maintain the configuration parameters of the cognitive bias function library, topological feature extraction algorithm, and convergence criterion logic unit.

Images