Sustainable consumption behavior simulation system fusing large language model and multi-agent

Abstract

The application relates to the technical field of artificial intelligence, in particular to a sustainable consumption behavior simulation system fusing a large language model and multiple intelligent agents, which comprises an external semantic generation module and an internal behavior evolution module; the external semantic generation module constructs multiple types of large language model intelligent agents, generates semantic intervention information according to state parameters of the internal behavior evolution module, and converts the semantic intervention information into standardized numerical signals through a preset mapping function; the internal behavior evolution module constructs consumer intelligent agents with double-layer attributes, adopts an optimized social influence model process based on an interaction network topology, completes cognitive interaction through social similarity calculation, uncertainty interval verification and nonlinear attitude updating, realizes prediction of cognitive behavior tendency through embedded neural networks, and simultaneously feeds back a group new state to the external module to form a closed loop. The system has both semantic dynamics and behavior simulation reality, and significantly improves the accuracy and practicability of sustainable consumption behavior evolution simulation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, in particular to a sustainable consumption behavior simulation system fusing large language model and multi-agent. BACKGROUND

[0002] Consumer sustainable consumption behavior is not a single static decision, but a dynamic evolution process under the coupling of individual cognition, social interaction, media guidance and brand communication and other multiple internal and external factors. Clarifying the formation and evolution mechanism of this behavior is the core prerequisite for optimizing brand strategy and realizing low-carbon transformation.

[0003] Traditional researches mostly rely on static statistical methods such as structural equation model and logistic regression, which can only depict the cross-sectional characteristics of consumer behavior in a specific situation, and cannot represent the dynamic evolution trajectory and group interaction effect of behavior. To break through this limitation, simulation technologies such as multi-agent (Agent) system (MAS) are introduced into the field of consumer behavior research and become the mainstream modeling tool. Through presetting fixed internal attributes (such as green consumption preference and conformity tendency) and static external environment rules (such as fixed policy influence coefficient and brand communication intensity) for consumer Agent, the technology simulates the interaction between individuals and the interaction between individuals and the environment, and then deduces the evolution law of behavior at the group level.

[0004] In the technical system of consumer cognitive communication simulation, the opinion dynamics model is the core supporting technology, and the Deffuant model is the classical technical scheme. Its core mechanism is that when the opinion difference between two consumer Agents is lower than the fixed confidence threshold, the two Agents will approach to form group consensus in a linear average way. To overcome the shortcomings of traditional opinion dynamics models such as Deffuant model in depicting social heterogeneity and cognitive evolution complexity, the SIM(Grow & Flache) model is proposed. Based on social identity theory and uncertainty theory, through social similarity calculation, dynamic uncertainty interval verification and asymmetric influence mechanism, the model can more realistically simulate complex social psychological processes such as attitude polarization and group differentiation of individuals in social networks, and its theoretical framework covers core technical modules such as social similarity calculation, uncertainty interval and influence condition definition, nonlinear cognitive updating mechanism and dynamic adjustment of uncertainty level.

[0005] In recent years, single Agent system (LLM-Agent) based on large language model (LLM) has been gradually applied to social behavior simulation field due to its strong semantic generation and environment perception ability. The technology can generate dynamic text intervention information through prompt and context, adjust the text content according to the simulation feedback, and convert the semantic information into numerical variables recognizable by the simulation system, which has significant technical advantages in social media communication and other scenarios.

[0006] However, the existing technical system still has many defects, which are difficult to meet the research needs of consumer sustainable consumption behavior simulation in complex scenarios, as follows:

[0007] First, the multi-agent system (MAS) technology has the dual defects of semantic loss and static modeling. On the one hand, external information such as news reports, policy documents, and brand narratives, which are rich in context and persuasive logic, are simplified into single fixed numerical parameters, losing semantic information highly relevant to consumer decision-making. On the other hand, it cannot adapt to the dynamic changes of media topic evolution, policy refinement, and brand strategy adjustment, resulting in low model restoration and large deviation between simulation results and actual consumer behavior.

[0008] Second, the traditional opinion dynamics model (represented by the Deffuant model) has the core problem of oversimplified cognitive interaction mechanism. First, it assumes that individual attitude updating has symmetry and singularity, ignoring the psychological uncertainty and the selective influence of social identity and circle recognition on information absorption. Second, the fixed confidence threshold setting cannot adapt to real-world phenomena such as opinion polarization and attitude opposition, making it difficult to reproduce complex social psychological processes such as group polarization and cognitive convergence.

[0009] Third, although the SIM(Grow & Flache) model theoretically improves the ability to depict the complexity of social attitude evolution, it has the key problem of high computational complexity, with a computational complexity of Direct application in large-scale multi-agent simulation scenarios will result in low computational efficiency and a sharp increase in simulation time, making it difficult to meet the real-time and scalability requirements of industry-level sustainable consumption behavior simulation.

[0010] Fourth, the single-agent system based on large language models (LLM-Agent) has the bottleneck of insufficient integration of LLM and multi-agent systems. First, the application scope is limited, and there is no systematic technical solution in the field of social economy and consumption behavior simulation. Second, there is a lack of a two-way feedback mechanism, and the semantic information generated by LLM-Agent cannot be quantified to affect group behavior evolution, and the dynamic state of consumers cannot guide LLM-Agent to optimize intervention information. Third, there are engineering defects such as high computational resource consumption, insufficient output stability, and poor repeatability, which make it difficult to meet the requirements of controllability, efficiency, and explainability of the simulation system, and the deep linkage between semantic information and behavior simulation has not been achieved.

[0011] In summary, the existing technology cannot build a simulation model that integrates multiple internal and external factors, accurately depicts the dynamic evolution of sustainable consumption behavior, adapts to complex social psychological mechanisms, and has reasonable computational cost. Building such a model has become a key technical requirement for promoting industrial green transformation and developing scientific low-carbon policies. SUMMARY

[0012] To this end, the present application aims to solve the technical problems that the traditional multi-agent sustainable consumption behavior simulation model has the problems of static external environment intervention, missing semantic information, simplified cognitive interaction mechanism, difficulty in simulating complex social phenomena such as attitude polarization, lack of system-level bidirectional closed-loop feedback in the fusion of large language models and multi-agent systems, limited application range, high operation cost and insufficient output stability, so as to provide a sustainable consumption behavior simulation system fusing large language models and multi-agent systems. The system comprises an external semantic generation module and an internal behavior evolution module;

[0013] The external semantic generation module is configured to: construct multi-type large language model agents, each type of large language model agent calling its corresponding external knowledge base based on the current state parameters generated by the internal behavior evolution module, dynamically generating unstructured semantic intervention information, and converting the semantic intervention information into standardized numerical signals that can be parsed by the internal behavior evolution module through a preset mapping function;

[0014] The internal behavior evolution module is configured to: construct a plurality of consumer agents and a regulatory agent deployed in an interaction network;

[0015] At each simulation step, the standardized numerical signals are received; based on an optimized social influence model, the plurality of consumer agents are driven to perform cognitive interaction within their network neighbor range, and the sustainable consumption cognitive value of each consumer agent is calculated and updated;

[0016] For each consumer agent, its updated sustainable consumption cognitive value, product feature signals and scene context signals parsed from external signals are assembled into an input vector, which is input into its embedded behavior decision neural network. After forward propagation and nonlinear mapping of the behavior decision neural network, the sustainable consumption behavior tendency value of the consumer agent at the next simulation step is output;

[0017] The regulatory agent monitors the latest state of all consumer agents in real time, calculates macro indicators at the group level, and encapsulates them as a new state vector, which is fed back to the external semantic generation module in real time through a preset data interface as input for generating the next round of semantic intervention information, thereby forming a closed-loop simulation loop.

[0018] In an embodiment of the present application, based on the optimized social influence model, the plurality of consumer agents are driven to perform cognitive interaction within their network neighbor range, and the sustainable consumption cognitive value of each consumer agent is calculated and updated, specifically including:

[0019] Based on the Watts-Strogatz small-world network model, an interactive topology environment is constructed for multiple consumer agents. By adjusting the average degree K and the reconnection probability β, it adapts to real social network simulation scenarios with different connection densities and randomness. The average degree K controls the initial number of connections for each agent, adjusting the overall connection density of the network. The reconnection probability β controls the reconstruction probability of long-range connections, introducing random connections while maintaining local clustering characteristics. In this interactive topology environment, the cognitive interactions between the multiple consumer agents are restricted to their direct network neighbors.

[0020] Based on the aforementioned interactive topology environment, at each simulation step, for any agent... With any of its neighboring intelligent agents Based on the difference in their perceived values ​​of sustainable consumption, the social similarity between the two groups was calculated. :

[0021] ,in, For intelligent agents Intelligent agents with neighbors The social similarity between them is used to quantify how close they are in their understanding of sustainable consumption; and respectively intelligent agents and Neighbor Intelligent Agent The sustainable consumption perception value is a real number that changes continuously within the interval [0,1].

[0022] Based on the aforementioned social similarity and intelligent agent The static attributes are used to perform acceptability checks on the influence of neighboring intelligent agents;

[0023] Update the agent based on the verified neighbor agents. Sustainable consumption awareness value.

[0024] In one embodiment of the present invention, based on the social similarity and the agent... The static attributes are used to perform acceptability checks on the influence of neighboring agents, specifically including:

[0025] Based on the preset similarity threshold Preliminary distinction of interaction types: when When, agents i and j are determined to be a cognitively similar group; when If so, they are identified as a group with cognitive differences;

[0026] Based on the aforementioned cognitively similar groups, for each consumer intelligent agent Set an internal acceptance threshold. By comparing the absolute values ​​of the cognitive differences between the two With the internal acceptance threshold Perform acceptance checks:

[0027] like Then determine the neighboring intelligent agent Cognition is at the level of the intelligent agent We accept this impact within an acceptable range;

[0028] Otherwise, determine the neighboring intelligent agent. Cognitive differences exceed those of intelligent agents Within acceptable limits, reject this impact;

[0029] Wherein, the internal acceptance threshold By intelligent agents It is calculated from static attributes including susceptibility and trust level.

[0030] In one embodiment of the present invention, the updating agent The formula for calculating the perceived value of sustainable consumption is as follows:

[0031] ,in, For intelligent agents At simulation time Sustainable consumption perception value; The preset learning rate parameter is used to control the overall rate of cognitive updates; For intelligent agents Its neighboring intelligent agents The connection weights between the links are used to characterize the importance of the connection to cognitive impact; For intelligent agents The set of neighbors in the interactive topology environment;

[0032] The connection weight Based on intelligent agents With any of its neighboring intelligent agents Social similarity calculated from the current perception value of sustainable consumption And / or the inherent properties of the connecting edge pre-assigned in the small-world network topology are dynamically determined.

[0033] In one embodiment of the present invention, the updating agent The formula for calculating the perceived value of sustainable consumption also includes a social stress item. It is used to quantify the conformity pressure an individual experiences due to their perception of the behavioral choices of the majority in their social circle; its intensity is related to the intelligence agent. a proportion of sustainable consumers in direct neighbor agents, wherein the sustainable consumer is defined as a neighbor agent whose sustainable consumption behavior tendency value exceeds a preset behavior adoption threshold;

[0034] introducing the social pressure term After that, the agent The calculation formula of the sustainable consumption cognition value of the agent is expanded as:

[0035] , wherein, is a social pressure influence coefficient.

[0036] In an embodiment of the present application, the behavior decision-making neural network is a BP neural network, and the structure thereof comprises:

[0037] An input layer comprising three neurons respectively receiving: the sustainable consumption cognition value of the agent at the current simulation time, a product sustainability intensity signal representing the environmental protection property of the product, and a product supply facility adaptability signal representing the purchase convenience and facility support;

[0038] A hidden layer configured with 5 to 64 neurons and adopting a ReLU activation function for nonlinear transformation;

[0039] An output layer comprising one neuron, which adopts a Sigmoid activation function to map the weighted sum of the neurons to the interval [0, 1] and output the sustainable consumption behavior tendency value at the next simulation time.

[0040] In an embodiment of the present application, the supervisory agent monitors the latest state of all consumer agents in real time, calculates the macro indicators at the group level, encapsulates them as a new state vector, and feeds them back to the external semantic generation module in real time through a preset data interface, specifically comprising:

[0041] At the end of each simulation step , the updated evolution states of all consumer agents are collected and aggregated in real time, and based on the updated evolution states, the macro indicators at the group level are calculated, including: the proportion of sustainable consumers, the group average sustainable consumption cognition and the group average sustainable consumption behavior;

[0042] The macro indicators are encapsulated as an actual state vector representing the actual state of the overall evolution of the group at the end of the current simulation step ;

[0043] The actual state vector is fed back to the external semantic generation module through a preset data interface, which is used to trigger the semantic intervention generation process of the external semantic generation module at the next simulation step . ​​

[0044] In an embodiment of the present application, the semantic intervention information is converted into a standardized numerical signal that can be parsed by the internal behavior evolution module through a preset mapping function, specifically including an inside-out mapping step and an outside-in mapping step:

[0045] In the inside-out mapping step, the current simulation step is received, and the current state vector representing the evolution state of the consumer agent group fed back by the internal behavior evolution module at the end ;

[0046] Based on the current state vector , a template-based prompt word construction function is used to generate semantic prompt words with targeted intervention goals for at least one type of external large language model agent ; the process is formally represented as: ;

[0047] In the outside-in mapping step, the semantic prompt words are input into the corresponding large language model agent to trigger the generation of structured semantic response information , which contains at least one quantifiable semantic influence parameter ;

[0048] Through a preset influence mapping matrix , at least one semantic influence parameter in the semantic response information is mapped to the expected influence amount on the consumer agent state vector : ;

[0049] Based on the expected influence amount and the current state vector , the target state vector that the system expects to achieve at the next simulation step is calculated, and the process is formally represented as: ; the target state vector is the output standardized numerical signal.

[0050] In an embodiment of the present application, the influence mapping matrix is configured to linearly combine the semantic influence parameters of different agents according to preset weights to calculate the direct influence amount on different dimensions of the state vector; its expression is as follows:

[0051] , where a response value representing a brand agent, a response value representing a rule agent, a response value representing a media agent; is a pre-calibrated influence coefficient.

[0052] In an embodiment of the present application, each of the consumer agents includes two types of attribute information independent of each other and associated with each other: one type is static attribute information for representing inherent social identity and underlying characteristics of the consumer, which remains stable throughout the simulation period and does not change dynamically with the interaction process, and includes herd mentality, trust, and willingness to accept sustainable consumption; the other type is dynamic attribute information for providing target evolution variables of cognitive interaction and consumption decision, which includes sustainable consumption cognition value and sustainable consumption behavior tendency value.

[0053] The above technical solution of the present application has the following beneficial effects compared with the prior art:

[0054] Firstly, by constructing an external semantic generation layer driven by a large language model, the static and numerical external intervention parameters in traditional simulation are converted into dynamic and semantic-rich intervention information flow, effectively solving the problems of semantic loss and staticity in external environment modeling, greatly improving the authenticity and ecological validity of the simulation environment, and accurately reflecting the dynamic evolution process of media, policy, and brand information;

[0055] Secondly, the optimized small-world network topology-based SIM (Grow&Flache) model (ESW-SIM) is introduced into the group behavior evolution layer, which innovatively uses optimized social similarity calculation and weighted average cognition updating mechanism, significantly improving the computational efficiency and large-scale simulation capability of the model, while preserving the core features of social dynamics such as attitude convergence and group differentiation, it can efficiently simulate thousands or even tens of thousands of consumer agents, providing a reliable computing tool and explanatory support for macro trend analysis and policy simulation of sustainable consumption behavior;

[0056] Thirdly, the built semantic-behavior double-layer closed-loop simulation framework realizes bidirectional feedback between the two layers through standardized interfaces and semantic-numerical mapping functions Φ, which not only allows external intervention to dynamically adjust according to the group evolution state, but also guarantees the computational efficiency and result stability of large-scale simulation, and strengthens the self-adaptive ability of the system;

[0057] ​​​Fourthly, through the cooperation of the three types of LLM-Agent, the SIM(Grow & Flache) model of the optimized embedded small-world network and the BP neural network, the complete path from semantic intervention to cognitive update to behavior decision-making is opened up, the cognitive-behavioral conversion path is accurately simulated, the simulation deviation of traditional research is reduced, the prediction ability of the adoption rate of sustainable consumption behavior is significantly improved, and the prediction data is highly consistent with the empirical data.

[0058] Fifthly, the technical scheme verifies good practicability in the clothing sustainable consumption scene, can truly present the characteristics of group cognitive convergence and attitude differentiation, has excellent field expandability, and can provide reliable simulation tools and decision support for policy making and brand strategy optimization. BRIEF DESCRIPTION OF DRAWINGS

[0059] In order to make the content of the present application easier to be clearly understood, the present application will be further described in detail below according to specific embodiments of the present application and in conjunction with the drawings.

[0060] Figure 1 is a sustainable consumption behavior simulation model structure diagram provided by an embodiment of the present application, which fuses a large language model and multiple agents;

[0061] Figure 2 is a LLM-Agent running mechanism diagram provided by an embodiment of the present application;

[0062] Figure 3 is a BP neural network structure model diagram for consumption behavior decision evolution provided by an embodiment of the present application;

[0063] Figure 4 is a dynamic representation model scene diagram of sustainable consumption behavior taking the clothing industry as an example. DETAILED DESCRIPTION

[0064] The present application will be further described below in conjunction with the drawings and specific embodiments, so that those skilled in the art can better understand the present application and implement it, but the embodiments are not limiting to the present application.

[0065] Referring to Figure 1 and Figure 2 , the present application provides a sustainable consumption behavior simulation system fusing a large language model and multiple agents, which comprises an external semantic generation module and an internal behavior evolution module.

[0066] The external semantic generation module is configured to: build a multi-type large language model agent, each type of large language model agent calls its corresponding external knowledge base based on the current state parameter generated by the internal behavior evolution module, dynamically generates unstructured semantic intervention information, and converts the semantic intervention information into standardized numerical signals that can be parsed by the internal behavior evolution module through a preset mapping function;

[0067] The internal behavior evolution module is configured to: build a plurality of consumer agents and regulatory agents deployed in a small-world network; at each simulation step, the internal state of the consumer agent is iteratively updated based on the standardized numerical signals; based on the updated agent state, the cognitive interaction, social communication and sustainable consumption decision process of the consumer group are simulated and deduced in multiple dimensions, and the new state parameters of the consumer group are fed back to the external semantic generation module to form a closed-loop simulation loop.

[0068] Further, the external semantic generation module is used to build three types of functional large language model agents (LLM-Agent), namely media agents (media Agent), rule agents (rule Agent) and brand agents (brand Agent), and the core function is to build a dynamic semantic intervention mechanism to generate targeted semantic information with scene adaptability, to accurately simulate media information diffusion, rule constraint intervention, brand dissemination guidance and other external social environmental elements, to accurately simulate the targeted intervention process of consumer sustainable consumption cognition formation and decision behavior evolution.

[0069] The three types of agents are respectively configured with four core functional sub-modules of environment perception, semantic planning, text generation and interface transmission; a supporting resource support system is simultaneously built, including a media dissemination knowledge base, an intervention rule database (with built-in reward-type and constraint-type intervention logic), and a brand sustainable attribute information base (covering product-side and supply-side green feature data), to provide basic resource support for the operation of each agent.

[0070] A product sustainability sub-agent and a facility supply sub-agent are additionally deployed for the characteristic requirements of the brand Agent: the product sustainability sub-agent stores product environmental material parameters, carbon footprint accounting data, green certification qualifications and other product-side core information; the facility supply sub-agent stores brand green supply chain architecture, recycling facility spatial distribution, and green purchase channel convenience, etc. Supply-side data, realize hierarchical representation and accurate call of brand-side sustainable attribute information.

[0071] A one-way data interaction interface is constructed for the environment perception sub-module of all Agents and the internal behavior evolution module, a standardized data receiving protocol is configured, and real-time collection of core simulation state parameters output by the internal behavior evolution module is ensured, including group average cognitive value, cognitive difference, behavior adoption rate, consumer price sensitivity, and environmental concern, etc. At the same time, a JSON standardized packaging protocol is configured for the interface transmission sub-module, fixed fields such as Agent type, intervention type, specific intervention semantic text, and generation timestamp are defined, and the standardization and analyzability of semantic information transmission are ensured.

[0072] The three types of LLM-Agent all follow a unified environment perception, semantic planning, text generation, and interface transmission four-core functional sub-module cooperative operation logic, and the specific execution steps are as follows:

[0073] The environment perception sub-module obtains the target state parameters of the current simulation period from the internal behavior evolution module through the preset data interaction interface, including the cognitive difference of the focused group of media Agents, the average cognitive level, the behavior adoption rate of the focused group of rule Agents, and the historical simulation feedback data, the price sensitivity of consumers focused by brand Agents, and the environmental concern, to provide accurate decision-making basis for semantic generation;

[0074] The semantic planning sub-module generates a targeted intervention semantic logic framework based on the collected state parameters: the media Agent calls the media transmission knowledge base to match the topic-oriented framework; the rule Agent calls the intervention rule database to match the reward or constraint rule framework; the brand Agent calls the brand sustainable attribute information base and generates the corresponding transmission content framework in parallel with the product / supply end sub-agent;

[0075] The text generation sub-module converts the semantic planning framework into a prompt word (Prompt) that can be parsed by a large language model, and calls the large language model to generate natural language semantic text that conforms to the intervention logic, including text context adaptability, emotional orientation, or persuasion features, such as "high cost-effective green clothing selection guide", "green consumption subsidy application specification", etc., avoiding single and numerical rigid expressions;

[0076] The interface transmission sub-module standardizes the generated natural language semantic text according to the preset JSON structure, and generates the final semantic intervention information.

[0077] The semantic intervention information is converted into a standardized numerical signal that can be parsed by the model algorithm in the internal behavior evolution module through a preset mapping function Φ (i.e. a double-layer mapping semantic intervention closed-loop mechanism, DM-CLoSeM), which includes an inward-outward mapping step and an outward-inward mapping step:

[0078] In the aforementioned inside-out mapping step (state awareness and prompt generation), the supervisory agent ( At each simulation step Real-time monitoring of consumer intelligent agents ( The evolutionary status of the group was analyzed, and key macroeconomic indicators were calculated, including the proportion of sustainable consumers. ( , For indicator functions, (Adoption threshold for consumer behavior) and average group perception of sustainable consumption. ( , Indicates the current simulation step size , No. (where N is the number of consumer agents) and the average sustainable consumption behavior of the group. ( , Indicates the current simulation step size , No. The sustainable consumption behavior of each consumer agent (where N is the number of consumer agents) is used to construct the current state vector. : ;

[0079] Based on the current state vector The three types of LLM-Agents—media, rules, and brand—are each constructed using a template-based prompt word function. Developing targeted semantic prompts: Media prompts Rule prompt words Brand prompts The inward-outward mapping process is formalized as follows:

[0080] ,in This is the vector for the prompt word.

[0081] In the outward-to-inward mapping step (semantic parsing and influence quantification), various LLM-Agents call large language models (such as the DeepSeek API) through encapsulated interfaces to generate structured JSON response information. Among them, the response information of the media agent Includes parameters such as conformity, trend, and propagation; and the response information of the rule agent. Includes parameters such as regulatory intensity, incentive intensity, and penalty intensity; brand agent response information. It includes parameters such as marketing intensity, sustainability, and commitment.

[0082] Specifically, , 、 The calculation of the three types of response values follows the unified technical process of "regulatory agent state summary→inward-outward mapping to generate prompt words→LLM call to generate structured JSON response→parse and write agent attributes→outward-inward mapping to achieve influence injection". Through the standardized process, the transformation from the system internal group state to quantifiable intervention parameters is realized. The specific calculation implementation process is as follows:

[0083] The calculation process is as follows:

[0084] When the system statistics get the sustainable consumer ratio is low, the media agent calls the template prompt word construction function to convert the macro indicators in into natural language prompt words , realizing the structured mapping from the system internal numerical state to semantic prompts. The formal representation of this process is: .

[0085] The prompt word template explicitly includes state description, parameter requirements, and format constraints, such as: "The current sustainable adoption rate is low, and the trend is rising. Please return JSON format data based on this state, including conformity, trend, and viral three parameters, with parameter value range [0, 1]". This ensures that the response generated by the large language model is targeted and analyzable.

[0086] The media agent calls the DeepSeek API (model parameter temperature=0.7) through the LLMClient tool, inputs the prompt word into the large language model, and triggers the generation of structured semantic response. An example of the structured JSON data returned by the large language model is as follows:

[0087] {"conformity":0.78,"trend":0.66,"viral":0.82}。

[0088] The system parses the JSON data generated by LLM, extracts the numerical values of conformity, trend, and viral fields, and performs [0, 1] interval range verification and standardization processing to ensure that the parameter values meet the system intervention requirements. Then, the parsed quantitative values are mapped to the internal attributes of the media agent: conformity corresponds to conformity influence strength , trend corresponds to trend amplification strength , the corresponding propagation coefficient , complete quantitative landing. At this time, the media response vector expression is: , combined with the above example data can be .

[0089] calculation process as follows:

[0090] rule agent calls the templated prompt construction function , the macro indicators in embedded in the preset prompt template, generating a targeted prompt with role positioning and format constraints , realize the structured transformation from the internal numerical state of the system to the semantic intervention demand, which is formally represented as .

[0091] Prompt templates explicitly include state descriptions, role instructions (rule makers), intervention targets, and output format requirements, such as: "The current sustainable adoption rate is 0.25, the group cognitive level is low and the differentiation is obvious. Please as the rule maker to give an intervention plan, return JSON format data, including regulation, incentive, and penalty three parameters, parameter value range is [0,1]", to ensure that the response generated by the large language model accurately matches the intervention needs.

[0092] The rule agent calls the DeepSeekAPI (model parameter temperature=0.7) through the LLMClient tool, inputs the prompt into the large language model, and triggers the generation of structured semantic responses related to regulation, incentive, and penalty. The structured JSON data returned by the large language model is as follows:

[0093] {"regulation":0.65,"incentive":0.80,"penalty":0.40}.

[0094] The system parses the JSON data generated by LLM, extracts the numerical values of the regulation, incentive, and penalty fields, and after range verification in the [0,1] interval, maps them to the core quantitative attributes of the rule agent, completing the quantification of .

[0095] Among them, the regulation field is mapped to the regulation intensity parameter , the incentive field is mapped to the incentive intensity parameter The penalty field is mapped to the penalty strength parameter ; at the same time, in order to ensure field compatibility and system scalability, the regulation agent additionally reserves synonymous internal fields: regulation_strength (equivalent to regulation_level, as an internal storage field of regulatory strength), and subsidy_level (equivalent to incentive_level, as an associated field of subsidy strength), the numerical consistency of which is strictly guaranteed through preset analysis mapping rules.

[0096] At this time, the rule response vector . Combining the above example data, we can get .

[0097] The calculation process is as follows:

[0098] The brand agent calls the templated prompt word construction function , which embeds the macro indicators in into the preset prompt word template to generate targeted prompt words with role positioning and format constraints . The prompt word template explicitly includes state description, professional role instructions (brand marketing and sustainable strategy expert), intervention needs, and output format requirements, such as: "The current sustainable consumer ratio is.25, and the group cognitive level is low. Please act as a brand marketing and sustainable strategy expert to develop an intervention plan, return JSON format data, including marketing (marketing strength), sustainability (sustainability focus), commitment (sustainable commitment), and price (price strategy) four parameters, parameter value range is [0,1]", to ensure that the response generated by the large language model accurately matches the professional scenario needs of brand intervention.

[0099] The brand agent calls the DeepSeekAPI (model parameter temperature=0.7) through the LLMClient tool, inputs the prompt word Ptbrand into the large language model, and triggers the generation of structured semantic responses related to brand marketing, sustainable value propagation, commitment endorsement, and price strategy. The structured JSON data returned by the large language model is as follows:

[0100] {"marketing":0.88,"sustainability":0.75,"commitment":0.90,"price":0.70}。

[0101] The system parses the JSON data generated by LLM, extracts the values ​​of four fields: marketing, sustainability, commitment, and price, and after validation within the range of [0,1], maps them to the core quantitative attributes of the brand agent, thus completing the quantitative implementation of Jtbrand. Specifically, the marketing field is mapped to the marketing intensity parameter. The sustainability field is mapped to a parameter indicating the degree of focus on sustainability information. The commitment field is mapped to a sustainability commitment credibility parameter. The price field is mapped to the price strategy strength parameter. At this point, the brand response vector .

[0102] LLM Influence Manager parses JSON response information And based on a preset influence mapping matrix Transform semantic parameters into expected impact on the consumer agent's state vector. : Influence mapping matrix Defined as:

[0103] ,in , , , These are pre-defined influence coefficients, corresponding to the influence intensity of brand marketing, regulatory rules, media conformity, and regulatory incentives, respectively.

[0104] Based on the expected impact and the current state vector The expected system size in the next simulation step is calculated. The target state vector achieved The process is formally represented as:

[0105] ,

[0106] The outward-to-inward mapping process is formalized as follows: ,in For the semantic generation function of a large language model, satisfying .

[0107] Closed-loop mechanism integration: Combining the two mappings above, the complete semantic intervention closed-loop mechanism is represented as a composite function: ,Right now: The new state vector The standardized numerical signal that can be parsed by the model algorithm in the internal behavior evolution module is used to drive the cognition and behavior evolution of the consumer agent at the next time.

[0108] It should be particularly noted that various LLM-Agent is only responsible for the dynamic generation of external semantic information, and does not participate in the direct interaction process between consumer agents; the whole process from semantic generation to state update is realized through the mapping function to realize end-to-end closed loop encapsulation, complete the complete mapping of "internal state perception→external semantic generation→influence force→internal state update".

[0109] The internal behavior evolution module aims to simulate the cognitive interaction and behavior decision-making process of consumers in a social network environment. This module contains two types of agents: consumer agents and regulatory agents.

[0110] A plurality of consumer agents are constructed, and each consumer agent is configured with two types of attribute information that are independent and related to each other:

[0111] One is static attribute information, which is used to represent the inherent social identity and underlying characteristics of the consumer individual, and remains stable throughout the simulation period and does not change dynamically with the interaction process. It includes herd mentality, trust, and initial willingness to accept sustainable consumption;

[0112] The other is dynamic attribute information, which is used to provide target evolution variables for cognitive interaction and consumption behavior decision-making, and is updated in real time with external intervention and group interaction. It includes sustainable consumption cognition (SCC) value and sustainable consumption behavior tendency (SCCB) value;

[0113] For regulatory agents, they are responsible for monitoring group state, and presetting to collect macro indicators such as average cognitive level, cognitive difference, behavior adoption rate and polarization index. They are configured with a communication interface with the external semantic generation module, and can pass the updated sustainable consumption cognition value to the external semantic generation module.

[0114] Based on the Watts-Strogatz small-world network model, the interaction topology of consumer agents is constructed, and two parameters, average degree K and reconnection probability β, are adjusted to control it. The characteristics of high clustering coefficient (simulating the local aggregation of the familiar circle) and short average path length (simulating the global accessibility of cross-circle information transmission) in the real social network are simulated. The interaction between agents is limited within their direct network neighbors, which not only conforms to the locality of real social interaction, but also greatly reduces the computational overhead of global matching in large-scale simulation.

[0115] The optimized SIM (Grow & Flache) model is used to realize the dynamic interaction and attitude update of sustainable consumption cognition among consumer agents, which is implemented in the following steps:

[0116] Initialize the basic dimensions of cognitive interaction for each consumer agent, including:

[0117] Sustainable consumption cognition value (SCC): represents the agent's cognitive level of sustainable consumption issues, with a value range of [0, 1], and the higher the value, the more positive the cognition. This value is the core variable in the simulation that evolves dynamically;

[0118] Static attribute set: including conformity (measuring the degree of individual influenced by group opinion), trust (affecting the credibility threshold of information acceptance), influence strength (the ability of individuals to output opinions externally), susceptibility (the sensitivity of individuals to external influence), etc. These attributes are set at the initialization of the simulation and remain relatively stable within the cycle, jointly regulating the strength and direction of social interaction of agents.

[0119] Based on the interaction topology environment, at each simulation step, for any agent and any neighbor agent , based on the difference in their sustainable consumption cognition values, the social similarity is calculated in real time using the following optimization formula:

[0120] where, is the social similarity between agent and neighbor agent , which quantifies the closeness of the two in sustainable consumption cognition; and are the sustainable consumption cognition values of agent and neighbor agent , respectively, with values as continuous real numbers within the interval [0, 1]. This formula represents the social distance in terms of the absolute difference in cognition values. The smaller the difference, the closer the social similarity to 1, indicating that the two cognition states are more similar, and theoretically the interaction willingness is stronger.

[0121] According to the calculated social similarity , determine the interaction tendency:

[0122] Set a similarity threshold (e.g. ) to preliminarily distinguish the interaction types:

[0123] When , it is determined that agent i and j are cognitively similar groups, and positive and consensus-promoting interactions are more likely to occur.

[0124] When , it is determined that the cognitive difference group, the interaction tendency is weakened, and even the rejection of views may be generated, which provides the basis for simulating group differentiation.

[0125] In the interaction process, not all the influence of the neighbors will be accepted. Through the positive interaction of the judgment, it still needs to be actually effective through the internal verification of the receiver. The system calculates an internal acceptance threshold according to the static attributes of the consumer agent i (especially the susceptibility and trust) . The core of this verification is to compare the cognitive difference and the individual tolerance:

[0126] If , it is determined that the cognition of the neighbor agent is within the acceptable range of the agent , and the influence is accepted;

[0127] Otherwise, it is determined that the cognitive difference of the neighbor agent exceeds the acceptable range of the agent , and the influence is rejected.

[0128] This optimization mechanism replaces the complex dynamic uncertainty interval verification in the original model with a clear and stable attribute-driven threshold , which ensures the rationality of the model behavior while significantly improving the computational efficiency.

[0129] For the neighbor agent that passes the above verification, the weighted average mechanism based on the local structure of the network is used to update the cognitive state of the agent :

[0130] , wherein is the sustainable consumption cognition value of the agent at the simulation moment , and is the updated sustainable consumption cognition value of the agent at the next simulation moment ; is the sustainable consumption cognition value of the neighbor agent at the simulation moment ; is the neighbor set of the agent in the small-world network; is the connection weight, which can be dynamically calculated according to the social similarity or the preset network edge strength, so as to give different neighbors different influence; is the global learning rate parameter, which is used to control the overall update speed and prevent iteration shock.

[0131] In the sustainable consumption cognition updating process of the consumer agent i, in order to more truly simulate the social effects such as herd mentality, a social pressure term can be additionally introduced in the above basic updating formula . The social pressure term simulates the herd mentality pressure generated by the individual due to the perceived behavior norms of the social circle in which it is located, and the strength value is usually positively correlated with the proportion of neighbors of the agent i who have been identified as sustainable consumers. Specifically, the neighbors whose sustainable consumption behavior tendency value exceeds the predefined behavior decision threshold are defined as sustainable consumers.

[0132] After introducing the social pressure term , the updating formula of the sustainable consumption cognition value can be extended as:

[0133] , wherein, is the social pressure influence coefficient. The extended formula drives the evolution of individual cognition by coupling the influence of neighbor opinion difference (main part of the formula) and the pressure of neighbor behavior consensus (social pressure term), thereby achieving comprehensive simulation of the "opinion exchange" and "behavior demonstration" mechanisms in the social influence process.

[0134] After the cognition update is completed, the sustainable consumption behavior tendency value of the consumer agent is predicted by a pre-trained BP neural network (as shown in Figure 3 ) embedded in it. Specifically, the structure of the BP neural network is as follows:

[0135] The input layer includes 3 neurons, which respectively receive: the sustainable consumption cognition value of the agent at the current simulation time , and the product feature signal (i.e., the product sustainability intensity signal PS representing the environmental properties of the product) and the scene context signal (i.e., the product supply facility adaptability signal PFSC representing the purchase convenience and facility support) obtained by analyzing the external signals; wherein the external product sustainability intensity signal PS is used to represent the strength of attributes such as brand environmental certification and carbon footprint; the external product supply facility adaptability signal PFSC is used to represent the convenience of recycling facilities, the price subsidy for purchasing green products, and other external contexts;

[0136] The hidden layer is configured with 5 to 64 neurons, which performs weighted summation operation on the above three-dimensional feature vector through learnable weight parameters, and then performs nonlinear mapping through the ReLU function to convert the original signal into high-order discriminative features, solve the gradient vanishing problem, and learn the complex interaction patterns between features;

[0137] An output layer comprising one neuron, which performs weighted aggregation on the high-order discriminative features, and then maps the aggregated result to a value in the interval [0, 1] through a Sigmoid activation function, and outputs a sustainable consumption behavior tendency value at the next simulation step When the predicted value is 1, it indicates that the sustainable consumption behavior tendency is extremely strong. When the predicted value is 0, it indicates that the sustainable consumption behavior tendency is extremely weak, and the tendency is traditional consumption.

[0138] At the end of each simulation step , the regulatory intelligent agent collects and aggregates the updated evolution states of all consumer intelligent agents in real time, and based on the updated evolution states, calculates macro indicators at the group level, including: group average sustainable consumption cognition value, sustainable consumer proportion, and behavior adoption speed; and encapsulates the macro indicators as an actual state vector representing the overall evolution result of the group at the end of the current simulation step . The actual state vector is fed back to the external semantic generation module through a preset data interface, to trigger the semantic intervention generation process of the external semantic generation module at the next simulation step , forming a closed-loop regulation and control of "cognitive evolution → state feedback → semantic intervention".

[0139] In the following, taking the clothing industry as a typical application scenario, the sustainable consumption behavior dynamic simulation scheme based on the large language model and the multi-agent system proposed in the present application is verified, and the specific implementation of the software and hardware architecture of the system is described in detail, to fully disclose the technical scheme of the present application, and to ensure that the present application can be reproduced by the person skilled in the art based on the present embodiment.

[0140] The simulation basic data sources of the present embodiment are divided into two categories, and both are subjected to a standardization preprocessing process: one is consumer shopping behavior basic data, which is derived from the consumer behavior and shopping habit dataset, which includes 3900 consumer records, and after being randomly reordered and expanded by 3 times, 11700 consumer records are obtained, covering 20 dimensions of behavior characteristics such as demographic characteristics, purchase history, brand preference, and price sensitivity, which are used to initialize the static attributes and consumption preferences of the consumer intelligent agent after data cleaning, outlier removal, and standardization processing; the other is consumer sustainable consumption cognition and attitude questionnaire data, which is collected through a professional online research platform, and 2000 valid questionnaires are obtained, which cover core psychological indicators such as sustainable clothing cognition level, environmental attitude, purchase willingness, and value identification, and after reliability and validity testing, they are used to set the initial cognitive state and decision parameters of the consumer intelligent agent. ​

[0141] To balance the computational efficiency and the observability of simulation results, 2000 consumer agents are configured in the embodiment; the social interaction topology of the consumer agents adopts the Watts-Strogatz small-world network structure, the average degree of the network is set to 6, and the reconnection probability is set to 0.1, so as to reproduce the characteristics of high clustering coefficient and short average path length of the real social network, and the initial sustainable consumption cognitive value of the consumer agent is set to be subject to the normal distribution with a mean of 0.25 and a standard deviation of 0.18, in combination with the actual statistical results of the questionnaire survey; a total of 50 time steps are set for the simulation period, so as to completely simulate the long-term evolution process of the consumer cognition and behavior under external intervention.

[0142] As shown in Figure 4 , the simulation process of the embodiment is divided into four core stages, which are executed in a closed loop in time sequence:

[0143] (1) System initialization stage: based on the preprocessed questionnaire data and consumer behavior data, the complete digital portrait of the consumer agent is constructed, specifically including static attributes (conformity, trust, etc.), initial cognitive state (sustainable consumption issue cognitive value) and network topology connection relationship, which lays a foundation for subsequent simulation.

[0144] (2) Cognitive interaction stage: the optimized SIM (Grow&Flache) model of the embedded small-world network (ESW-SIM) is executed at each time step, and the consumer agent sequentially completes social similarity calculation, influence receiving condition verification, weighted average cognitive updating and other operations, so as to realize the cognitive propagation and attitude evolution within the group.

[0145] (3) External semantic intervention stage: the intervention logic of the LLM agent is triggered every 5 time steps, wherein the media agent dynamically generates controversial or incentive sustainable clothing propagation content according to the group cognitive difference and the average cognitive level; the rule agent outputs differentiated policy intervention schemes such as subsidies and authentication based on the reinforcement learning algorithm; the brand agent (including product sustainability sub-agent and facility supply sub-agent) adaptively generates "economic" or "environmental" brand narrative content according to the consumer price sensitivity and environmental concern, and all intervention contents are converted into quantitative signals through the semantic-numerical mapping interface and input into the simulation system.

[0146] (4) Behavior decision-making stage: the consumer agent inputs the iteratively updated sustainable consumption cognitive value, combined with the external environmental signals such as product sustainability and supply facility conditions, into the embedded BP neural network to complete decision mapping; the BP neural network is a three-layer structure, the input dimension is 3, the number of hidden layer nodes is 32, and the output layer adopts the Sigmoid activation function, and finally outputs the sustainable clothing purchase behavior tendency value in the interval of 0-1.

[0147] After 50 time steps of full simulation, the system reaches a stable evolution state, and the core indicators show significant optimization: the sustainable clothing adoption rate increases to 67.3%, an increase of 42.3 percentage points from the initial baseline; the average sustainable consumption awareness of consumers increases from 0.31 to 0.72; the effects of different types of external intervention can be quantified: the price subsidy policy increases the purchase conversion rate by 18.5%, the media content increases the awareness propagation speed by 32.7%, and the brand sustainability narrative increases the product trust by 27.2%; based on the above simulation results, the following optimization suggestions are output for the clothing industry: on the product side, focus on improving clothing durability and optimizing pricing rationality; on the communication side, use empirical data to convey the long-term value of sustainable clothing; on the policy side, design a tiered subsidy program to improve the efficiency of fund utilization; on the channel side, optimize the shopping experience to enhance the conversion efficiency of awareness to behavior.

[0148] The system can use a high-performance computing architecture to meet the computing needs of large-scale social simulation. The high-performance computing architecture usually includes multi-core processors, large-capacity memories, high-performance GPU acceleration units, hierarchical storage systems, and high-speed network interconnections. In a specific embodiment, the computing node is configured with a 48-core AMD EPYC 9654 CPU and 128GB of memory, the acceleration computing unit uses a NVIDIA RTX 6000 GPU (96GB of video memory), the storage system uses a hierarchical scheme combining high-speed solid state drives and large-capacity mechanical hard drives, and the network uses high-speed Ethernet. This configuration can support the concurrent operation of more than 10,000 agents, meeting the computing needs of neural network inference and LLM service calls.

[0149] The software system uses a modular Python architecture, and each functional module interacts and cooperates through Python standard interfaces and function calls.

[0150] The system core consists of the following types of functional modules:

[0151] Agent module: contains five types of agents, including consumers, social media, policies, brands, and regulators, implemented as independent Python classes. Each type of agent encapsulates its specific behavior logic and state management functions;

[0152] Simulation model module: contains multiple independent simulation model Python files. Among them, a main simulation model file is responsible for coordinating the global process and calling other models; there are multiple model files that implement network evolution and cognitive interaction logic based on different theories (such as SIM and Deffuant);

[0153] Tool module: contains several independent Python tool files, which respectively implement AI large language model API calling (through HTTP request), configuration file management (reading JSON format, etc.), data export (generating CSV files, etc.), and other auxiliary functions;

[0154] User interface module: contains Python files of graphical interface based on Qt framework, which provides windows for parameter input, simulation control and result display;

[0155] Experiment and visualization module: contains scripts for running comparative experiments, parameter tuning experiments, and Python code for drawing charts. The system implements key processes in the following ways:

[0156] Data and configuration: initial data and parameters are obtained by reading external CSV, JSON, etc. files.

[0157] Simulation execution: the main program calls the update method and model calculation function of each type of agent in a step-by-step loop; external semantic integration: when needed, send an HTTP request to the API endpoint of a large language model such as DeepSeek to obtain a text response, and parse it into a signal usable by the system; result output: record the state data during the simulation process in memory, and finally write it to a CSV file or display it through a chart.

[0158] The system runs in a standard Python environment and relies on scientific computing and network analysis libraries such as NumPy and PyTorch. Its architecture allows the addition of agent types, agent behaviors, replacement of network models, or integration of new AI large language model services by modifying or replacing specific Python class and function files.

[0159] The present application is described with reference to flowcharts and / or block diagrams according to the embodiments of the present application; it should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions; these computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The device for implementing the functions specified in one or more flows and / or blocks Figure 1 The device for implementing the functions specified in one or more flows and / or blocks

[0160] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flow Figure 1 one or more flows and / or blocks Figure 1 one or more blocks or multiple blocks to perform a function specified in the flow or flows and / or blocks.

[0161] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flow Figure 1 one or more flows and / or blocks Figure 1 one or more blocks or multiple blocks to perform a function specified in the flow or flows and / or blocks.

[0162] Obviously, the above-mentioned embodiments are only examples for clearly illustrating the present application, and are not intended to limit the present application; on the basis of the above-mentioned description, other different forms of changes or variations can also be made by those skilled in the art. Here, all the embodiments are not required to be enumerated, and the obvious changes or variations derived therefrom are still within the protection scope of the present application.

Claims

1. A simulation system for sustainable consumer behavior that integrates a large language model and multi-agent systems, characterized by: External semantic generation module and internal behavior evolution module; The external semantic generation module is configured to: construct multiple types of large language model intelligent agents; each type of large language model intelligent agent, based on the current state parameters generated by the internal behavior evolution module, calls its corresponding external knowledge base to dynamically generate unstructured semantic intervention information; and converts the semantic intervention information into standardized numerical signals that can be parsed by the internal behavior evolution module through a preset mapping function. The internal behavior evolution module is configured to: construct multiple consumer agents and regulatory agents deployed in the interactive network; At each simulation step, the standardized numerical signal is received; based on the optimized social impact model, multiple consumer agents are driven to perform cognitive interactions within their network neighbors, and the sustainable consumption cognitive value of each consumer agent is calculated and updated. For each consumer agent, its updated sustainable consumption cognitive value is assembled together with the product feature signal and scenario context signal obtained from external signals into an input vector, which is then input into its embedded behavioral decision neural network. After forward propagation and nonlinear mapping of the behavioral decision neural network, its sustainable consumption behavior tendency value in the next simulation step is output. The regulatory agent monitors the latest status of all consumer agents in real time, calculates macro-level indicators at the group level, encapsulates them into new state vectors, and feeds them back to the external semantic generation module in real time through a preset data interface as input for generating the next round of semantic intervention information, thereby forming a closed-loop simulation circuit.

2. The sustainable consumption behavior simulation system integrating a large language model and multi-agent technology as described in claim 1, characterized in that, Based on the optimized social impact model, multiple consumer agents are driven to engage in cognitive interactions within their network neighbors, and the sustainable consumption cognitive value of each consumer agent is calculated and updated, specifically including: Based on the Watts-Strogatz small-world network model, an interactive topology environment is constructed for multiple consumer agents. By adjusting the average degree K and the reconnection probability β, it adapts to real social network simulation scenarios with different connection densities and randomness. The average degree K controls the initial number of connections for each agent, adjusting the overall connection density of the network. The reconnection probability β controls the reconstruction probability of long-range connections, introducing random connections while maintaining local clustering characteristics. In this interactive topology environment, the cognitive interactions between the multiple consumer agents are restricted to their direct network neighbors. Based on the aforementioned interactive topology environment, at each simulation step, for any agent... With any of its neighboring intelligent agents Based on the difference in their perceived values ​​of sustainable consumption, the social similarity between the two groups was calculated. : ， in, For intelligent agents Intelligent agents with neighbors The social similarity between them is used to quantify how close they are in their understanding of sustainable consumption; and respectively intelligent agents and Neighbor Intelligent Agent The sustainable consumption perception value is a real number that changes continuously within the interval [0,1]. Based on the aforementioned social similarity and intelligent agent The static attributes are used to perform acceptability checks on the influence of neighboring intelligent agents; Update the agent based on the verified neighbor agents. Sustainable consumption awareness value.

3. The sustainable consumption behavior simulation system integrating a large language model and multi-agent technology as described in claim 2, characterized in that, Based on the aforementioned social similarity and intelligent agent The static attributes are used to perform acceptability checks on the influence of neighboring agents, specifically including: Based on the preset similarity threshold Preliminary distinction of interaction types: when When, agents i and j are determined to be a cognitively similar group; when If so, they are identified as a group with cognitive differences; Based on the aforementioned cognitively similar groups, for each consumer intelligent agent Set an internal acceptance threshold. By comparing the absolute values ​​of the cognitive differences between the two With the internal acceptance threshold Perform acceptance checks: like Then determine the neighboring intelligent agent Cognition is at the level of the intelligent agent We accept this impact within an acceptable range; Otherwise, determine the neighboring intelligent agent. Cognitive differences exceed those of intelligent agents Within acceptable limits, reject this impact; Wherein, the internal acceptance threshold By intelligent agents It is calculated from static attributes including susceptibility and trust level.

4. The sustainable consumption behavior simulation system integrating a large language model and multi-agent technology as described in claim 2, characterized in that, The updated intelligent agent The formula for calculating the perceived value of sustainable consumption is as follows: ,in, For intelligent agents At simulation time Sustainable consumption perception value For intelligent agents At the next simulation moment Updated awareness of sustainable consumption; The preset learning rate parameter is used to control the overall rate of cognitive updates; For intelligent agents Its neighboring intelligent agents The connection weights between the links are used to characterize the importance of the connection to cognitive impact; For intelligent agents The set of neighbors in the interactive topology environment; The connection weight Based on intelligent agents With any of its neighboring intelligent agents Social similarity calculated from the current perception value of sustainable consumption And / or the inherent properties of the connecting edge pre-assigned in the small-world network topology are dynamically determined.

5. The sustainable consumption behavior simulation system integrating a large language model and multi-agent technology as described in claim 4, characterized in that, The updated intelligent agent The formula for calculating the perceived value of sustainable consumption also includes a social stress item. It is used to quantify the conformity pressure an individual experiences due to their perception of the behavioral choices of the majority in their social circle; its intensity is related to the intelligence agent. The proportion of sustainable consumers among the direct neighbor agents is positively correlated, wherein the sustainable consumer is defined as a neighbor agent whose sustainable consumption behavior tendency value exceeds a preset behavior adoption threshold; Introducing the aforementioned social pressure item Then, update the intelligent agent. The formula for calculating the perceived value of sustainable consumption is extended as follows: ,in, This represents the social pressure influence coefficient.

6. The sustainable consumption behavior simulation system integrating a large language model and multi-agent technology as described in claim 1, characterized in that, The behavioral decision neural network is a BP neural network, and its structure includes: The input layer contains three neurons, which respectively receive: the agent's sustainable consumption cognition value at the current simulation moment, the product sustainability strength signal representing the product's environmental attributes, and the product supply facility adaptability signal representing purchase convenience and facility support. The hidden layer contains 5 to 64 neurons and uses the ReLU activation function for nonlinear transformation. The output layer contains one neuron. The sigmoid activation function is used to map the weighted sum of the neurons to the interval [0,1], and outputs the sustainable consumption behavior tendency value at the next simulation time.

7. The sustainable consumption behavior simulation system integrating a large language model and multi-agent technology as described in claim 1, characterized in that, The regulatory agent monitors the latest state of all consumer agents in real time, calculates macro-level indicators at the group level, encapsulates them into new state vectors, and feeds them back to the external semantic generation module in real time through a preset data interface. Specifically, this includes: At each simulation step At the end, the updated evolution states of all consumer agents are collected and aggregated in real time, and based on the updated evolution states, macro indicators at the group level are calculated, including: the proportion of sustainable consumers, the average sustainable consumption awareness of the group, and the average sustainable consumption behavior of the group. The macroscopic indicator is encapsulated into a representation of the current simulation step size. At the end, the actual state vector of the overall population evolution result ; The actual state vector is transmitted through a preset data interface. Feedback is sent to the external semantic generation module to trigger the external semantic generation module in the next simulation step. The semantic intervention generation process.

8. The sustainable consumption behavior simulation system integrating a large language model and multi-agent technology as described in claim 1, characterized in that, The semantic intervention information is transformed into a standardized numerical signal that can be parsed by the internal behavior evolution module through a preset mapping function, specifically including an inside-out mapping step and an outside-in mapping step: In the inside-out mapping step, the current simulation step size is received. At the end, the current state vector, which represents the evolutionary state of the consumer agent group, is fed back by the internal behavior evolution module. ; Based on the current state vector It uses a template-based prompt word construction function. Generate semantic cue words with targeted intervention objectives for at least one type of external large language model agent. The formal representation of this process is as follows: ; In the outward-to-inward mapping step, the semantic prompt words are... The input is fed into the corresponding large language model agent, triggering it to generate structured semantic response information. The semantic response information It includes at least one quantifiable semantic impact parameter; Through a pre-set influence mapping matrix The semantic response information At least one semantic influence parameter is mapped to the expected influence on the consumer agent's state vector. : ; Based on the expected impact and the current state vector The expected system size in the next simulation step is calculated. The target state vector achieved The process is formally represented as: The target state vector This is the output standardized numerical signal.

9. The sustainable consumption behavior simulation system integrating a large language model and multi-agent technology as described in claim 8, characterized in that, The influence mapping matrix It is configured to linearly combine the semantic influence parameters of different agents according to preset weights to calculate the direct influence on different dimensions of the state vector; its expression is as follows: ， in, This represents the response value of the brand's intelligent agent. This represents the response value of the rule-based agent. This represents the response value of the media agent; , , , The influence coefficient is a pre-calibrated value.

10. The sustainable consumption behavior simulation system integrating a large language model and multi-agent technology as described in claim 1, characterized in that, Each consumer agent includes two independent but related types of attribute information: one type is static attribute information, which is used to characterize the consumer's inherent social identity and underlying traits. It remains stable throughout the simulation period and does not change dynamically with the interaction process. It includes conformity, trust, and willingness to accept sustainable consumption; the other type is dynamic attribute information, which is used to provide target evolution variables for cognitive interaction and consumption decision-making. It includes sustainable consumption cognitive value and sustainable consumption behavior tendency value.

Images