Decision-making method and system based on social dynamics evolution and hybrid reinforcement calibration

Abstract

This invention discloses a decision-making method and system based on social dynamics evolution and hybrid reinforcement calibration. It generates a massive number of heterogeneous intelligent agents with independent cognitive architectures, drives their "perception-cognition-decision" cycle using a large language model, introduces a scale-free network-based social dynamics model to simulate the nonlinear propagation of information and opinion polarization in strong and weak relationship networks, and uses a Bayesian structural equation model with sparse anchor data to perform posterior calibration of the simulation system. This invention can perform minute-level and full-scale rehearsals for macro-strategy deployment or micro-target testing, accurately capturing emergent group behavior.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and social computing, and more specifically, to a decision-making method and system based on social dynamics evolution and hybrid enhancement calibration. Background Technology

[0002] In public affairs management and business operations, decision-makers often face an "information fog." Traditional research methods (such as questionnaires and in-person focus groups) have serious limitations:

[0003] 1. Sample bias and "spiral of silence": When faced with sensitive topics or under group pressure, real individuals often hide their true intentions (Hawthorne effect), causing data to fail to reflect real behavior.

[0004] 2. Lack of dynamic evolution prediction: Traditional static snapshots cannot predict how a strategy will go through the entire process of "fermentation-outbreak-subsidence" on social networks after its release, especially failing to capture "viral spread" or "public opinion backlash" on platforms like short videos.

[0005] 3. High time and trial-and-error costs: Whether it is launching a new product or piloting a policy, the real-world cost of trial and error is extremely high and irreversible.

[0006] While existing large-scale AI models possess strong text generation capabilities, they lack the topological structure of social interactions and cannot simulate the mutual influence between individuals. Although similar advanced technologies internationally (such as agent-based simulation systems) have achieved multi-agent parallelism, their model parameters are mostly derived from abstractions of behavioral patterns under specific cultural backgrounds (such as individualism). Therefore, when applied to high-density social networks that emphasize collective interaction (such as acquaintances who value "relationships"), their parameter assumptions and interaction mechanisms often suffer from insufficient adaptability. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a decision-making method and system based on social dynamics evolution and hybrid reinforcement calibration to construct a "digital parallel society" that can not only simulate individual micro-decision-making, but also focuses on simulating macro-emergence at the group level.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A decision-making method based on social dynamics evolution and hybrid enhancement calibration includes the following steps:

[0010] Step S1: Based on multi-source desensitized statistical data, use generative adversarial networks to generate a set of intelligent agents with multi-level cognitive features and dynamic memory;

[0011] Step S2: Construct a virtual environment that includes physical constraints and information transmission media, and transform the decision scheme to be tested into an environmental stimulus signal with multi-dimensional attributes;

[0012] Step S3: Drive the agent to perform a two-level decision-making process of "inner loop - utility evaluation" and "outer loop - social game", and comprehensively calculate the agent's response probability and emotional polarity to the incentive signal.

[0013] Step S4: Based on the small-world network topology, simulate the cascading propagation of opinions among groups, the echo chamber effect, and group polarization to generate a dynamic public opinion evolution map.

[0014] Step S5: Using small sample feedback data from the real world as anchor points, Bayesian inference is used to correct the cognitive parameters and interaction weights of the agent, and the final decision risk assessment report is output.

[0015] Furthermore, in step S1, the set of agents includes demographic attributes, psychological characteristics, value vectors, and historical memory maps.

[0016] Furthermore, in step S2, the excitation signal includes attribute dimensions, cost constraints, and propagation modes.

[0017] Furthermore, in step S3, the agent is modeled using a utility function that includes a social modification term. For excitation signal Willingness to accept :

[0018]

[0019]

[0020] In the formula, For agents based on personal preferences and attribute dimensions in the excitation signal The intrinsic utility; The conformity coefficient; For intelligent agents The social neighborhood; For intelligent agents In neighboring nodes Influence weight; Excitation signal In neighboring nodes The signal status; It is random noise; For intelligent agents For excitation signal Overall utility; It is a natural constant.

[0021] Furthermore, in step S4, a cultural resonance factor is introduced. Simulating explosive spread in a high-density social network, the view update equation is:

[0022]

[0023] In the formula, For intelligent agents At any moment Opinion value; This refers to the assimilation rate; Based on the current topic's popularity The nonlinear gain function; For intelligent agents The social neighborhood; Excitation signal At any moment The corresponding social consensus attribute value.

[0024] Furthermore, in step S5, a Bayesian structural equation model is used for posterior calibration, with the objective function being to minimize the simulation output distribution. With anchor point distribution KL divergence between:

[0025]

[0026] In the formula, These are the system parameters to be corrected. These are the calibrated system parameters; For observed variables; The Kullback-Leibler divergence; For anchor point distribution in the real world Under the condition, observed variables The true conditional probability distribution; To the current simulation system parameters and its output distribution Under the condition of observation variables The simulated conditional probability distribution.

[0027] This invention also provides a decision-making system based on social dynamics evolution and hybrid enhancement calibration, the system comprising:

[0028] A swarm generator is used to instantiate intelligent agents with long short-term memory and thought chain reasoning capabilities;

[0029] The strategy simulation engine includes distributed computing units that support large-scale concurrent reasoning, used to perform dual-loop cognitive reasoning and social dynamics evolution calculations, simulating the decision-making process and opinion propagation of a group under incentive signals;

[0030] The virtual-reality alignment module is used to access small sample feedback data from the real world in real time, and correct simulation parameters through Bayesian inference and backpropagation to achieve alignment and calibration between the simulation system and real social dynamics.

[0031] A multi-dimensional insight dashboard is used to visualize group sentiment heatmaps, decision-making conversion funnels, public opinion evolution trends, and potential risk nodes.

[0032] The beneficial effects of this invention are:

[0033] 1. This invention constructs a dual-loop cognitive model that integrates individual intrinsic utility and social network influence, and introduces a cultural resonance factor to simulate the nonlinear characteristics of information dissemination, enabling the system to realistically reproduce the evolution, polarization, and outbreak of viewpoints in real society.

[0034] 2. This invention conducts full-scale testing in a completely virtual "digital parallel society," avoiding the high economic and social costs associated with conducting large-scale pilot projects or questionnaire surveys in the real world. Simultaneously, the system requires only a minimal amount of publicly available anchor data for calibration, eliminating the need to collect sensitive individual information, thus achieving high-precision simulation with "zero privacy risks." This provides a safe and low-cost pre-testing environment for policy evaluation and business strategy testing.

[0035] 3. This invention enables the system to flexibly adapt to simulation requirements of different social network densities and group interaction patterns by designing adjustable parameters such as conformity coefficient, influence weight, and cultural resonance factor.

[0036] 4. This invention is based on a hybrid enhancement calibration method that combines Bayesian structure equations and KL divergence minimization. This method enables the system to continuously correct model parameters and reduce biases caused by large model illusions by utilizing a small amount of feedback data continuously generated in the real world. This forms a continuous learning closed loop that aligns the simulation system with real social dynamics, maintaining the reliability and timeliness of prediction results. Attached Figure Description

[0037] Figure 1 This is a flowchart of a decision-making method based on social dynamics evolution and hybrid enhancement calibration in this embodiment;

[0038] Figure 2 This is a flowchart of a dual-loop cognitive reasoning method in this embodiment;

[0039] Figure 3 This is a structural framework diagram of a decision-making system based on social dynamics evolution and hybrid enhancement calibration in this embodiment. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Example: A decision-making method based on social dynamics evolution and hybrid reinforcement calibration addresses the time lag, Hawthorne effect, and data silos inherent in existing survey methods. It constructs a digital twin social environment statistically isomorphic to the real world. By generating a massive number of heterogeneous intelligent agents with independent cognitive architectures, it drives their "perception-cognition-decision" cycle using a large language model. A scale-free network-based social dynamics model is introduced to simulate the nonlinear propagation of information and opinion polarization in strong and weak relationship networks. Finally, a Bayesian structural equation model is used to perform posterior calibration of the simulation system using sparse anchor data. This method can perform minute-level and full-scale rehearsals of macro-strategy deployment or micro-target testing, accurately capturing emergent group behavior.

[0042] Specifically, such as Figure 1 As shown, it includes the following steps;

[0043] Step S1: Construct a high-dimensional heterogeneous virtual group: Based on multi-source desensitized statistical data (including but not limited to demographic yearbooks, consumer behavior data, social survey data, etc.), use generative adversarial networks to generate a set of intelligent agents with multi-level cognitive characteristics and dynamic memory. Each intelligent agent is given a unique cognitive embedding vector, including demographic attributes, psychological characteristics, value vectors and historical memory maps.

[0044] In this step, adversarial training of generative adversarial networks ensures that the statistical distribution of the generated set of agents is consistent with that of the target real population, while maintaining reasonable heterogeneity among individuals.

[0045] Step S2, Define the dynamic interaction field: Construct a virtual environment that includes physical constraints (such as geographic space and resource distribution) and information dissemination media (such as social networks and media channels), and transform the decision scheme to be tested into an environmental stimulus signal with multi-dimensional attributes. This signal includes the following dimensions:

[0046] Attribute Dimensions Used to describe excitation signals Its core features include policy content, product functions, and price;

[0047] Cost constraints Used to describe individual adoption incentive signals The costs involved, such as money, time, and social costs;

[0048] Propagation mode This is used to describe the channels and methods of signal propagation, such as official announcements, social media pushes, word-of-mouth marketing, etc.

[0049] In this step, a virtual environment is constructed to provide perceptual input to the intelligent agent and define its possible action space and interaction rules.

[0050] Step S3: Perform dual-loop cognitive reasoning: drive the agent to perform a two-layer decision-making process of "inner loop - utility evaluation" and "outer loop - social game", and comprehensively calculate the agent's response probability and emotional polarity to the incentive signal.

[0051] like Figure 2 As shown, for each agent, the two-layer cognitive process of executing "inner loop - utility evaluation" and "outer loop - social game" is as follows:

[0052] Inner loop - utility evaluation, the agent bases its decisions on its own preferences. and the attribute dimensions of the excitation signal Calculate intrinsic utility ;

[0053] Outer-loop social game: The agent observes the reactions of other agents in its social neighborhood and dynamically adjusts its own decision-making tendencies based on social influence.

[0054] intelligent agent For excitation signal Overall utility Calculated by the following formula:

[0055]

[0056] In the formula, total utility Used to measure intelligent agents After considering individual intrinsic preferences and surrounding social influences, the motivational signal The overall tendency; The conformity coefficient reflects the strength of collectivism in the environment and ranges from 0 to 1. For intelligent agents The social neighborhood; For intelligent agents In neighboring nodes Influence weight; Excitation signal In neighboring nodes The signal status is set to 1 if accepted and -1 if rejected. The noise is random, simulating irrational fluctuations in individuals.

[0057] Ultimately, the intelligent agent Regarding the excitation signal Willingness to accept It is calculated using the following logical function:

[0058]

[0059] In the formula, It is a natural constant used to convert utility values. It is mapped to a probability value between 0 and 1.

[0060] Step S4, Evolutionary Social Dynamics Propagation: Based on the small-world network topology, simulate the cascading propagation of opinions among groups, the echo chamber effect, and group polarization to generate a dynamic public opinion evolution map.

[0061] Furthermore, by introducing cultural resonance factors This simulates the explosive propagation phenomenon in high-density social networks. The viewpoint update equation for each agent is as follows:

[0062]

[0063] In the formula, For intelligent agents At any moment Opinion value; Assimilation rate; controlling the speed at which opinions are influenced by others. Based on the current topic's popularity The nonlinear gain function, when the heat exceeds the threshold At that time, cultural resonance factors It grows exponentially, simulating the phenomenon of "trending searches" or "viral posts"; Excitation signal At any moment The corresponding social consensus attribute value.

[0064] This cultural resonance equation enables the system to simulate complex social phenomena such as echo chamber effect, group polarization, and opinion differentiation, generating a dynamic public opinion evolution map.

[0065] Step S5, Hybrid Enhancement Calibration: Using small sample feedback data from the real world as anchors, Bayesian inference is used to correct the cognitive parameters and interaction weights of the agent, and the final decision risk assessment report is output.

[0066] Specifically, a Bayesian structural equation model (BSEM) is used for posterior calibration, with the objective function being to minimize the simulation output distribution. With anchor point distribution KL (Kullback-Leibler) divergence between them:

[0067]

[0068] In the formula, The system parameters to be corrected include the agent's cognitive parameters (such as conformity coefficient). Personal preferences ) and interaction weights; These are the calibrated system parameters; For observed variables, such as decision conversion rate, group sentiment polarity, and public opinion tendency; Kullback-Leibler divergence is used to quantify the distance between the simulated conditional probability distribution and the true conditional probability distribution. For anchor point distribution in the real world Under the condition, observed variables The true conditional probability distribution; To the current simulation system parameters and its output distribution Under the condition of observation variables The simulated conditional probability distribution.

[0069] Updating system parameters using Markov chain Monte Carlo (MCMC) sampling method The system gradually corrects systematic biases caused by the illusion of a large model, aligning simulation results with real-world observation data. After calibration, a final decision risk assessment report is output, including predictions of group adoption rates, public opinion evolution trends, and potential risk points.

[0070] This embodiment also provides a decision-making system based on social dynamics evolution and hybrid enhancement calibration, such as... Figure 3 As shown, the system includes a swarm generator, a strategy simulation engine, a virtual-reality alignment module, and a multi-dimensional insight dashboard.

[0071] Among them, the group generator is responsible for instantiating intelligent agents with long short-term memory (LSTM) and thought chain (CoT) reasoning capabilities, and generating heterogeneous groups with multi-level cognitive characteristics and dynamic memory based on statistical data;

[0072] The strategy simulation engine is a distributed computing unit that supports large-scale concurrent reasoning. It is used to perform dual-loop cognitive reasoning and social dynamics evolution calculations to simulate the decision-making process and opinion propagation of a group under incentive signals.

[0073] The virtual-reality alignment module is used to access small sample feedback data from the real world in real time, and correct simulation parameters through Bayesian inference and backpropagation to achieve alignment and calibration between the simulation system and real social dynamics.

[0074] The multidimensional insight dashboard provides a visual interactive interface for visually displaying group sentiment heatmaps, decision-making conversion funnels, public opinion evolution trends, and potential risk nodes, supporting decision-makers in dynamic monitoring and strategy optimization.

[0075] The specific implementation plan for the above decision-making method is as follows:

[0076] 1. Instantiation of Heterogeneous Intelligent Agents

[0077] During system initialization, a large-scale virtual agent swarm (e.g., N=100,000) is generated based on multi-source data such as statistical yearbooks of the target region. Each agent... Endowed with unique cognitive embedding vectors ,include:

[0078] Basic attributes: age, occupation, disposable income, education level, etc.;

[0079] Psychological mapping: Generating personality scores (extraversion, agreeableness, conscientiousness, neuroticism, openness) using the Big Five personality scale, and setting risk aversion coefficients, etc.

[0080] Network location: Indicators such as degree centrality and betweenness centrality in virtual social networks determine its influence in information dissemination;

[0081] Historical memory map: Records the agent's past behavioral choices, social interactions, information access, etc., stored in the form of a long short-term memory (LSTM) network, and influences future decisions.

[0082] The attribute distribution of all agents is consistent with real statistical data, ensuring the representativeness of the virtual group.

[0083] 2. Simulation Cycles and Social Interaction

[0084] The simulation proceeds in discrete time steps. At each time step... The system broadcasts decision signals to the virtual environment. (For example, if a new regulation or new product is released), the intelligent agent will perform the following actions:

[0085] Independent thinking: Each agent uses a large language model (LLM) to parse the signal content and calculates the initial response probability based on its own preferences;

[0086] Social interaction: The agent observes the reactions of its neighboring nodes, especially the attitudes of high-influence nodes (opinion leaders). If there are strong emotional expressions among the neighbors, the agent will adjust its own attitude.

[0087] Opinion Update: Update the opinion value of each agent according to the cultural resonance equation to simulate the cascading effect of opinion propagation;

[0088] Behavioral output: Based on the final effect value, decide whether to adopt the incentive and update its historical memory.

[0089] This cycle continues for multiple time steps until the system reaches a stable state or completes the preset period.

[0090] 3. Bayesian calibration closed loop

[0091] After the initial simulation results are generated, the system connects to real-world "probe data" (i.e., anchor data), such as: the actual conversion rate in small-scale pilot areas, the true sentiment value of early public opinion on social media, and publicly available consumer behavior survey data.

[0092] The calibration process is as follows:

[0093] Difference Calculation: Calculating Simulation Distribution Distribution in reality The statistical differences between them were measured using KL divergence as a measure of difference.

[0094] Parameter tuning: Adjust the agent's general parameters, such as price sensitivity, through Markov chain Monte Carlo (MCMC) sampling or variational inference. Conformity coefficient wait;

[0095] Iterative optimization: Repeat the simulation-calibration cycle until the difference between the simulation results and the real data converges to a preset threshold. within;

[0096] Report generation: Outputs a calibrated simulation report, including visual analyses such as group behavior prediction, risk heatmap, and decision transformation funnel.

[0097] Through this closed-loop calibration mechanism, the system can achieve high-fidelity simulation using only a small amount of publicly available data while protecting user privacy, providing decision-makers with minute-level and full-scale strategy simulation capabilities.

[0098] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A decision-making method based on social dynamics evolution and hybrid enhancement calibration, characterized in that, Includes the following steps: Step S1: Based on multi-source desensitized statistical data, use generative adversarial networks to generate a set of intelligent agents with multi-level cognitive features and dynamic memory; Step S2: Construct a virtual environment that includes physical constraints and information transmission media, and transform the decision scheme to be tested into an environmental stimulus signal with multi-dimensional attributes; Step S3: Drive the agent to execute a two-level decision-making process of "inner loop - utility evaluation" and "outer loop - social game", and comprehensively calculate the agent's response probability and emotional polarity to the incentive signal. Step S4: Based on the small-world network topology, simulate the cascading propagation of opinions among groups, the echo chamber effect, and group polarization to generate a dynamic public opinion evolution map. Step S5: Using small sample feedback data from the real world as anchor points, Bayesian inference is used to correct the cognitive parameters and interaction weights of the agent, and the final decision risk assessment report is output.

2. The decision-making method based on social dynamics evolution and hybrid enhancement calibration according to claim 1, characterized in that, In step S1, the set of agents includes demographic attributes, psychological characteristics, value vectors, and historical memory maps.

3. The decision-making method based on social dynamics evolution and hybrid enhancement calibration according to claim 1, characterized in that, In step S2, the excitation signal includes attribute dimensions, cost constraints, and propagation modes.

4. The decision-making method based on social dynamics evolution and hybrid enhancement calibration according to claim 1, characterized in that, In step S3, the agent is modeled using a utility function that includes a social modification term. For excitation signal Willingness to accept : In the formula, For agents based on personal preferences and attribute dimensions in the excitation signal The intrinsic utility; The conformity coefficient; For intelligent agents The social neighborhood; For intelligent agents In neighboring nodes Influence weight; Excitation signal In neighboring nodes The signal status; It is random noise; For intelligent agents For excitation signal Overall utility; It is a natural constant.

5. The decision-making method based on social dynamics evolution and hybrid enhancement calibration according to claim 1, characterized in that, In step S4, a cultural resonance factor is introduced. Simulating explosive spread in a high-density social network, the view update equation is: In the formula, For intelligent agents At any moment Opinion value; This refers to the assimilation rate; Based on the current topic's popularity The nonlinear gain function; For intelligent agents The social neighborhood; Excitation signal At any moment The corresponding social consensus attribute value.

6. The decision-making method based on social dynamics evolution and hybrid enhancement calibration according to claim 1, characterized in that, In step S5, a Bayesian structural equation model is used for posterior calibration, with the objective function being to minimize the simulation output distribution. With anchor point distribution KL divergence between: In the formula, These are the system parameters to be corrected. These are the calibrated system parameters; For observed variables; The Kullback-Leibler divergence; For anchor point distribution in the real world Under the condition, observed variables The true conditional probability distribution; To the current simulation system parameters and its output distribution Under the condition of observation variables The simulated conditional probability distribution.

7. A decision-making system based on social dynamics evolution and hybrid enhancement calibration for implementing the method of claim 1, characterized in that, The system includes: A swarm generator is used to instantiate intelligent agents with long short-term memory and reasoning capabilities. The strategy simulation engine includes distributed computing units that support large-scale concurrent reasoning, used to perform dual-loop cognitive reasoning and social dynamics evolution calculations, simulating the decision-making process and opinion propagation of a group under incentive signals; The virtual-reality alignment module is used to access small sample feedback data from the real world in real time, and correct simulation parameters through Bayesian inference and backpropagation to achieve alignment and calibration between the simulation system and real social dynamics. A multi-dimensional insight dashboard is used to visualize group sentiment heatmaps, decision-making conversion funnels, public opinion evolution trends, and potential risk nodes.

Images