A low-frequency decision dynamic simulation method based on a large language model

Abstract

The application discloses a low-frequency decision dynamic simulation method based on a large language model, and aims at the problems of cognitive rigidity, decision frequency distortion and waste of computing resources when the existing fixed step simulation framework processes 'high cost, low frequency' decisions. The application constructs a dynamic decision interval mechanism: through initial delay allocation, the heterogeneity of agents is realized, external environmental disturbance and internal cognitive evolution are dynamically calculated in real time to determine the decision trigger point, and a storage module based on natural language reflection is used, so that the agent can self-correct the decision interval according to the historical refusal reason, thereby realizing high-fidelity simulation of decision inertia. One typical scene of the method is household photovoltaic adoption behavior prediction and policy simulation, and the method can be popularized to long-term low-frequency decision scenes such as house purchase and education investment, so that the simulation precision is improved and the computing power consumption is greatly reduced.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and computational social science, and in particular to a method for dynamic simulation of low-frequency decision-making based on a large language model. Background Technology

[0002] With the rapid development of artificial intelligence technology, Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language processing, logical reasoning, and complex decision-making simulation. In recent years, researchers have begun to explore using agents driven by large models to simulate complex socioeconomic behaviors. However, existing simulation frameworks face significant technical bottlenecks when dealing with decision-making scenarios characterized by low frequency, high cost, and high perceived risk (such as residential solar PV adoption, home purchase, and long-term education investment).

[0003] In the field of behavioral simulation, the mainstream technical approaches can be broadly divided into two categories: Firstly, there is traditional agent-based modeling (ABM) based on rules or mathematical formulas. Early research (such as the study by Tesfatsion et al. in 2006) mainly relied on hard-coded logic or simple assumptions about maximizing economic utility. While these methods can provide some interpretability, they are essentially static rule bases, making it difficult to capture the true psychological hesitations of decision-makers when faced with complex policies and ambiguous public opinion. Especially when dealing with low-frequency decisions that are not entirely rational, traditional ABM often simplifies decision-making to a single threshold judgment (such as whether the investment payback period is less than a certain length), ignoring the cognitive evolution in human behavior.

[0004] Secondly, there are large-scale model simulation frameworks based on fixed step sizes. Current LLM agent simulation frameworks mainly focus on high-frequency, continuous, and low-weight decision-making scenarios. In the field of energy management, existing technologies (such as Chetty et al.'s research in 2025) simulate household electricity consumption behavior through large models. Since electricity consumption is continuous data generated by residents' daily lives, such simulations usually adopt a fixed-step decision-making mechanism, calling the large model in its entirety in each simulation cycle (such as hourly or daily) to generate the agent's electricity consumption behavior. In the field of social behavior simulation, there are frameworks such as the AgentSociety framework developed by the Center for Urban Science and Computing at Tsinghua University, and the YuLan-OneSim simulation system proposed by Ren et al. These frameworks are dedicated to characterizing the social, work, and consumption activities of agents within a 24-hour period. Their core logic lies in maintaining a tight time rhythm, enabling agents to remain active and dynamically adjust their behavior under the drive of needs and cognition.

[0005] However, in real life and socio-economic systems, there are numerous low-frequency, high-cost, and long-cycle decision-making scenarios. These include not only decisions about installing distributed photovoltaic systems, but also decisions about homeownership, purchasing large-scale household energy storage equipment, long-term education investment planning for children, and large-scale equipment upgrades for businesses. These decisions are characterized by extremely long decision windows, extremely low trigger frequencies, and are often accompanied by long observation and hesitation periods.

[0006] While LLM has brought significant changes to behavioral simulation, its application in the energy sector still has a clear gap. The main gap lies in the misalignment of decision-making attributes: existing research largely focuses on high-frequency, low-weight daily behaviors (such as daily electricity consumption), and has not been effectively extended to low-frequency, high-cost, and high-perceived-risk investment decisions, such as photovoltaic installations. Current LLM multi-agent framework simulations in the energy sector generally employ a full-volume periodic call strategy (i.e., querying all agents monthly), which is unsuitable for simulating low-frequency social adoption behaviors and can easily lead to misjudgments of diffusion rates.

[0007] It should be noted that the information disclosed in the background section above is only for understanding the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0008] The main objective of this invention is to overcome the deficiencies in the aforementioned background technology and provide a low-frequency decision dynamic simulation method based on a large language model.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A dynamic simulation method for low-frequency decision-making based on a large language model is used to simulate the diffusion trend of low-frequency, high-cost decision-making behavior, including the following steps: S1. Construct a heterogeneous group of intelligent agents and assign an initial decision delay to each agent to simulate the difference in the time of the first consideration of the target decision by the individual at the beginning of the simulation. S2. Monitor external environmental disturbances and internal cognitive states in real time, dynamically calculate the decision interval of each agent, and trigger explicit reasoning of the agent based on the decision interval or external disturbances, so that the agent only enters the decision state when the decision interval is zero or when it is subjected to a major external disturbance. S3. For the triggered agent, call the large language model to perform decision reasoning. If the decision result is rejection, generate a structured rejection reason and store it in the memory module. In subsequent decision cycles, the agent retrieves and reflects on historical rejection reasons to ensure the consistency of long-term decision logic. S4. Summarize the decision results of all agents and generate simulation output to predict the diffusion trend of low-frequency, high-cost behaviors.

[0010] Furthermore, in step S1, the allocation of the initial decision delay is based on the innovation diffusion theory combined with individual characteristics: the basic delay is generated by random distribution and adjusted according to the difference of each agent's environmental awareness relative to the preset benchmark level. The higher the environmental awareness, the shorter the initial delay, so as to simulate the heterogeneity of individuals' attention time to new technologies.

[0011] Furthermore, in step S2, the dynamic calculation of the decision interval for each agent includes: based on a preset baseline decision interval, and taking into account the agent's economic attributes, risk preferences, social structural constraints, and information acquisition capabilities, the decision interval value is obtained by weighted summation; wherein, factors such as high income, low risk aversion, property support, and multiple information channels are assigned weights to shorten the decision interval, and vice versa.

[0012] Furthermore, in step S2, the explicit reasoning that triggers the agent includes: automatic activation when the decision interval counter returns to zero, or immediate activation when a major disturbance event occurs in the external environment, so that the agent enters the decision state.

[0013] Furthermore, in step S3, generating structured rejection reasons and storing them in the memory module includes: the activated agent calling the large language model to perform chained reasoning; if a rejection decision is generated, the agent guides the model to output a formatted rejection reason and stores it in long-term memory; in subsequent decision cycles, the agent retrieves historical rejection reasons from the memory module and uses them as context input to the large language model to reflect on and maintain decision consistency.

[0014] Furthermore, in step S3, after reflecting on the historical reasons for rejection, the reflected information is converted into a numerical cognitive inertia parameter. This parameter is used to adjust the length of the next decision interval, so that the decision frequency of the agent is affected by historical cognitive feedback.

[0015] Furthermore, in step S3, after generating the structured rejection reasons, the next decision interval is fine-tuned based on the semantic content of the rejection reasons: if the rejection reason points to high cost, the decision interval is extended to simulate risk-averse waiting behavior; if the rejection reason points to insufficient information, the decision interval is shortened to simulate information search motivation.

[0016] Furthermore, in step S4, the simulation output includes generating a technology adoption diffusion curve, and by comparing the curve changes under different policy interventions, the incentive effect of the policy on low-frequency decision-making behavior is evaluated.

[0017] Furthermore, after step S1 or during the simulation, a step of dynamically introducing new agents is also included: according to the technology diffusion curve model and combined with random perturbations, the number of new agents entering the decision-making process is generated monthly to simulate the continuous emergence of potential adopters in the real market and avoid simulation distortion caused by sample aging.

[0018] Furthermore, the method can be applied to the accurate prediction and policy simulation scenarios of residential photovoltaic adoption behavior to simulate the photovoltaic system installation decision-making process of residents when facing electricity price fluctuations, subsidy policy adjustments, and neighborhood installation effects; or, the method can be applied to the home purchase decision simulation scenario to simulate the timing and hesitation behavior of families when facing changes in credit policies, regional development plans, and market expectation disturbances; or, the method can be applied to the education investment decision simulation scenario to simulate the decision delay and final choice behavior of families when considering their children's long-term education investment, influenced by policy guidance, social comparison, and information access capabilities.

[0019] The present invention has the following beneficial effects: This invention proposes a dynamic simulation method for low-frequency decision-making based on a large language model. It constructs a Dynamic Decision Interval (DDI) mechanism to address the inherent shortcomings of existing large-scale model simulation frameworks in handling "high-cost, low-frequency" decisions such as residential photovoltaic adoption. Traditional fixed-step triggering mechanisms not only cause "cognitive rigidity" in agents due to high-frequency invalid queries, preventing them from responding correctly when real opportunities arise, but also cause dynamic distortion of simulation time by ignoring real-world "decision inertia" and "observation periods." Furthermore, in city simulations on a scale of tens of thousands, this leads to numerous invalid API calls and wasted computing power. To address these issues, this invention achieves agent heterogeneity through initial delay allocation and couples external policy disturbances with internal cognitive evolution to dynamically calculate decision trigger points. Simultaneously, it utilizes a storage module based on natural language reflection, enabling agents to self-correct decision intervals based on historical rejection reasons (such as cost or insufficient information), thereby achieving a high-fidelity simulation of human decision inertia.

[0020] The technical solution of this invention can be accurately applied to the prediction and policy simulation of residential photovoltaic adoption behavior, and can be extended to long-term, low-frequency decision-making scenarios such as home purchase and education investment. Compared with existing technologies, this invention has the following significant advantages: First, by using the DDI mechanism, it breaks through the unrealistic assumption of the traditional "all-person decision-making" approach, activating explicit reasoning only when external disturbances or internal cognitive triggers are perceived. This allows the active distribution of the agent group to accurately correspond in time and space to the real population "generating decision-making intentions," thereby realistically reproducing the nonlinear dynamic process of alternating "latency period" and "outbreak period" in the diffusion of social technology. Second, this on-demand activation logic greatly reduces the consumption of computing resources in large-scale simulations, significantly reducing redundant LLM calls under the same population size, making refined behavioral simulations at the city level and for tens of millions of people possible with limited resources. Finally, by combining DDI with a reflection-based storage module, this invention effectively solves the problems of agent state transitions and memory loss in simulation cycles lasting several years. The dynamic interval provides a reasonable evolutionary time for semantic correction, ensuring that the psychological evolution trajectory of the agent highly matches the prudent decision-making characteristics of real humans.

[0021] Other beneficial effects of the embodiments of the present invention will be further described below. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the structure of an embodiment of the present invention.

[0023] Figure 2 This is a framework diagram of a low-frequency decision dynamic simulation method based on a large language model, according to an embodiment of the present invention.

[0024] Figure 3 This is an evolution diagram of potential household adopters in the large-scale intelligent agent simulation of an embodiment of the present invention.

[0025] Figure 4 This is a comparison chart of photovoltaic installations in a certain city from 2020 to 2025 (simulated values ​​compared with actual values).

[0026] Figure 5 This is a comparison chart of installation volume under a fixed-step update mechanism. Detailed Implementation

[0027] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and not intended to limit the scope and application of the present invention.

[0028] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0029] This invention aims to address the problems of cognitive rigidity, decision frequency distortion, and wasted computational resources caused by the fixed-step-size triggering mechanism in existing large-scale simulation frameworks when dealing with "high-cost, low-frequency" decisions. It proposes a dynamic simulation method for low-frequency decisions based on Dynamic Decision Interval (DDI). This method constructs heterogeneous agents and allocates initial delays, dynamically calculating decision trigger points by combining external policy disturbances with internal cognitive evolution. Simultaneously, it utilizes a natural language-based reflection-based storage module to enable agents to self-correct decision intervals based on historical rejection reasons, thereby achieving high-fidelity simulation of decision inertia. This overcomes the unrealistic assumption of traditional "all-personnel decision-making," achieving precise screening and mapping of the decision-making population, significantly reducing the computational consumption of large-scale simulations, and significantly enhancing the cognitive consistency of agents in long-term simulations.

[0030] See Figure 1 and Figure 2 This invention provides a method for dynamic simulation of low-frequency decision-making based on a large language model, used to simulate the diffusion trend of low-frequency, high-cost decision-making behavior, including the following steps: Step S1: Construct a heterogeneous group of agents and assign an initial decision delay to each agent to simulate the difference in the time it takes for individuals to first consider the target decision at the beginning of the simulation.

[0031] In some embodiments, in step S1, the allocation of the initial decision delay is based on the innovation diffusion theory combined with individual characteristics: the base delay is generated by random distribution and adjusted according to the difference of each agent's environmental awareness relative to a preset baseline level. The higher the environmental awareness, the shorter the initial delay, so as to simulate the heterogeneity of individuals' attention time to new technologies.

[0032] Step S2: Monitor external environmental disturbances and internal cognitive states in real time, dynamically calculate the decision interval of each agent, and trigger explicit reasoning of the agent based on the decision interval or external disturbances, so that the agent only enters the decision state when the decision interval is zero or when it is subjected to a major external disturbance.

[0033] In some embodiments, step S2, the dynamic calculation of the decision interval for each agent includes: based on a preset baseline decision interval, and taking into account the agent's economic attributes, risk preferences, social structural constraints, and information acquisition capabilities, obtaining the decision interval value through a weighted summation; wherein, factors such as high income, low risk aversion, property support, and multiple information channels are assigned weights to shorten the decision interval, and vice versa.

[0034] In some embodiments, in step S2, the explicit reasoning of the agent includes: automatically activating when the decision interval counter returns to zero, or immediately activating when a major disturbance event occurs in the external environment, so that the agent enters the decision state.

[0035] Step S3: For the triggered agent, call the large language model to perform decision reasoning. If the decision result is rejection, generate a structured rejection reason and store it in the memory module. In subsequent decision cycles, the agent retrieves and reflects on historical rejection reasons to ensure the consistency of long-term decision logic.

[0036] In some embodiments, step S3, generating structured rejection reasons and storing them in the memory module, includes: the activated agent calling the large language model to perform chained reasoning; if a rejection decision is generated, the agent guides the model to output a formatted rejection reason and stores it in long-term memory; in subsequent decision cycles, the agent retrieves historical rejection reasons from the memory module and inputs them as context into the large language model to reflect on and maintain decision consistency.

[0037] In some embodiments, in step S3, after reflecting on the historical reasons for rejection, the reflected information is converted into a numerical cognitive inertia parameter. This parameter is used to adjust the length of the next decision interval, so that the agent's decision frequency is affected by historical cognitive feedback.

[0038] In some embodiments, step S3, after generating a structured rejection reason, further includes a step of fine-tuning the next decision interval based on the semantic content of the rejection reason: if the rejection reason points to high cost, the decision interval is extended to simulate risk-averse waiting behavior; if the rejection reason points to insufficient information, the decision interval is shortened to simulate information search motivation.

[0039] Step S4: Summarize the decision results of all agents and generate simulation output to predict the diffusion trend of low-frequency, high-cost behaviors.

[0040] In some embodiments, in step S4, the simulation output includes generating a technology adoption diffusion curve, and by comparing the curve changes under different policy interventions, the incentive effect of the policy on low-frequency decision-making behavior is evaluated.

[0041] In some embodiments, after step S1 or during the simulation, a step of dynamically introducing new agents is also included: according to the technology diffusion curve model and combined with random perturbations, the number of new agents entering the decision-making process is generated monthly to simulate the continuous emergence of potential adopters in the real market and avoid simulation distortion caused by sample aging.

[0042] In different embodiments, the method of the present invention can be applied to the simulation and policy evaluation of typical low-frequency, high-cost decision-making scenarios such as residential photovoltaic adoption, home purchase decisions, and education investment planning. It can also be extended to other socio-economic behavior simulation fields with characteristics of "high cost, low frequency, and long cycle," such as enterprise equipment upgrades, large-scale consumer goods purchases, and career transition choices. However, the present invention is not limited to these. Any improved or extended application based on a dynamic decision interval mechanism to simulate the decision-making process of individuals or groups in low-frequency, high-perceived-risk situations, as long as its core method utilizes initial delay allocation, external disturbance and cognitive evolution coupling triggering, and semantic feedback based on historical reflection to correct the decision interval, falls within the scope of the present invention.

[0043] In one specific embodiment, the method is used for accurate prediction of residential solar PV adoption behavior and policy simulation scenarios, specifically including the following steps: constructing a heterogeneous group of agents and assigning an initial decision delay to each agent to simulate the difference in the initial consideration time of individuals' solar PV installation decisions at the beginning of the simulation in a residential solar PV adoption scenario; monitoring external environmental disturbances and internal cognitive states in real time, dynamically calculating the decision interval of each agent, and triggering explicit reasoning of the agents based on the decision interval or external disturbances, so that the agents only enter the solar PV installation decision state when the decision interval is zero or when subjected to significant external disturbances; for the triggered agents, calling a large language model to perform solar PV installation decision reasoning, and if the decision result is rejection, generating structured rejection reasons and storing them in the memory module; in subsequent decision cycles, the agents retrieve and reflect on historical rejection reasons to ensure the consistency of long-term decision logic; summarizing the solar PV installation decision results of all agents to generate simulation output, which is used to predict the diffusion trend of residential solar PV adoption behavior and evaluate the policy incentive effect.

[0044] In another specific embodiment, the method is used in a home purchase decision simulation scenario, including the following steps: constructing a heterogeneous group of intelligent agents and assigning an initial decision delay to each agent to simulate the difference in the initial consideration time of a family's home purchase decision at the beginning of the simulation; monitoring external environmental disturbances and internal cognitive states in real time, dynamically calculating the decision interval of each agent, and triggering explicit reasoning of the agents based on the decision interval or external disturbances, so that the agents only enter the home purchase decision state when the decision interval is zero or when subjected to significant external disturbances; for the triggered agents, calling a large language model to perform home purchase decision reasoning, and if the decision result is rejection, generating structured rejection reasons and storing them in the memory module; in subsequent decision cycles, the agents retrieve and reflect on historical rejection reasons to ensure the consistency of long-term decision logic; summarizing the home purchase decision results of all agents to generate simulation output, which is used to predict the diffusion trend of home purchase behavior and evaluate the effect of policy regulation.

[0045] In another specific embodiment, the method is used in an education investment decision-making simulation scenario, including the following steps: constructing a heterogeneous group of intelligent agents and assigning an initial decision delay to each agent to simulate the difference in the initial consideration time of families' long-term education investment decisions for their children in the initial stage of the simulation scenario; monitoring external environmental disturbances and internal cognitive states in real time, dynamically calculating the decision interval of each agent, and triggering explicit reasoning of the agents based on the decision interval or external disturbances, so that the agents only enter the education investment decision state when the decision interval is zero or when subjected to significant external disturbances; for the triggered agents, calling a large language model to perform education investment decision reasoning, and if the decision result is rejection, generating structured rejection reasons and storing them in the memory module; in subsequent decision cycles, the agents retrieve and reflect on historical rejection reasons to ensure the consistency of long-term decision logic; summarizing the education investment decision results of all agents and generating simulation output for predicting the diffusion trend of education investment behavior and assessing the impact of relevant policies.

[0046] This invention proposes a dynamic simulation method for low-frequency decision-making based on a large language model. By constructing a dynamic decision-making interval mechanism, it fundamentally breaks through the unrealistic assumption of "fixed-step-size decision-making by all participants" in traditional simulations. Explicit reasoning by the corresponding agent is activated only when external environmental disturbances or internal cognitive evolution reach a trigger point. This allows the active distribution of the agent group to accurately map the real population "generating decision-making intentions" in the spatiotemporal dimension. It solves the problems of cognitive rigidity, decision frequency distortion, and misjudgment of macroscopic diffusion rates caused by high-frequency invalid inquiries in existing technologies when dealing with low-frequency, high-cost decisions. Furthermore, this on-demand activation logic significantly reduces the consumption of computing resources in large-scale social simulations, making refined behavioral simulations at the city level or even with tens of millions of people possible with limited resources. Simultaneously, by combining the dynamic decision-making interval with a storage module based on natural language reflection, this invention effectively maintains the cognitive consistency of the agents over simulation cycles lasting several years, ensuring that their psychological evolution trajectory highly matches the prudent decision-making characteristics of real humans. This provides high-fidelity simulation capabilities and reliable policy evaluation tools for low-frequency, high-cost decision-making scenarios such as residential photovoltaic adoption, home purchase, and education investment.

[0047] The following further describes specific embodiments of the present invention, algorithm examples, and experimental verification.

[0048] A dynamic simulation method for low-frequency decision-making based on a large language model addresses the problems of cognitive rigidity, decision frequency distortion, and wasted computational resources caused by the fixed-step triggering mechanism in existing large-model simulation frameworks when dealing with "high-cost, low-frequency" decisions such as residential photovoltaic adoption. The core of this method lies in constructing a Dynamic Decision Interval (DDI) mechanism. Specifically, this method achieves high-fidelity simulation of low-frequency decision-making processes through the following main steps:

[0049] First, through heterogeneous agent initialization and initial delay allocation, the system constructs heterogeneous agents with multidimensional attributes (such as economic status, environmental awareness, and housing type) based on statistical data of the target region. Building upon this, and combining innovation diffusion theory with individual characteristics, an initial decision delay is assigned to each agent, thus establishing a differentiated foundation for the simulation's starting point.

[0050] Secondly, during simulation, the system monitors the external environment (such as policy changes and electricity price adjustments) and internal cognitive states (such as the accumulation of neighborhood effects) in real time, and calculates the dynamic decision interval (DDI) of each agent through a weighted function. The system only activates the explicit inference module of the agent when the interval counter reaches zero or is triggered by a major external disturbance, thereby fundamentally solving the problem of decision frequency distortion caused by the traditional fixed step size method.

[0051] Once the agent is activated, the system performs cognitive consistency processing based on reflection. The activated agent invokes a large language model to make decisions. If a rejection occurs, the system guides the model to generate structured reasons for rejection and store them in the memory module. In subsequent decision cycles, the agent needs to retrieve and reflect on historical reasons to ensure the consistency of long-term decision logic.

[0052] As a preferred embodiment of the present invention, the system can also perform semantic-driven fine-tuning of the decision interval. This step automatically and positively corrects (extends) the interval of the next decision by parsing semantic information in the reflection module (e.g., "rejected due to high installation costs"), so as to refine the simulation of "risk-averse waiting" behavior in reality, thereby further enhancing the simulation depth of human decision-making inertia.

[0053] Finally, by combining simulation results with policy evaluation, the system summarizes the adoption decisions of all agents at each point in time, generates technology diffusion curves, and achieves high-fidelity simulation of low-frequency, high-cost investment behavior and prediction of policy responses.

[0054] In the above steps, heterogeneous agent initialization, dynamic decision interval calculation and activation triggering, and reflection-based cognitive consistency processing together constitute the core means of realizing the low-frequency decision simulation of this invention. Semantic-driven fine-tuning of the decision interval is a preferred solution for further improving simulation accuracy.

[0055] The following example simulates a household's decision on whether to install solar panels: Unlike existing multi-agent simulation studies based on large language models (LLM) that generally adopt the "full-scale periodic call" strategy (such as Chetty et al., 2025 simulation of daily electricity consumption behavior), this embodiment proposes a dynamic decision interval (DDI) mechanism for the typical low-frequency, high-cost, and high-cognitive-load technology adoption behavior of rooftop distributed photovoltaic systems.

[0056] The core idea behind this mechanism is that not all households are in an "active decision-making state" at every time step. In reality, most residents don't suddenly consider whether to install solar panels—a real decision-making process only begins when an individual actually encounters relevant information (such as policy announcements, neighbor installations, or company promotions) or perceives significant changes in the external environment (such as rising electricity prices or subsidy reductions). In other words, the LLM agent invoked by the model does not represent the entire population, but rather precisely corresponds to those real individuals who are "truly considering whether or not to install" at a specific moment. This design abandons the unrealistic assumption of "everyone making decisions monthly," instead focusing on real decision-making windows triggered by information or driven by context.

[0057] This mechanism not only significantly improves computational efficiency (avoiding tens of thousands of meaningless LLM calls), but more importantly, it significantly enhances the model's ecological validity—that is, the consistency between simulation results and real human behavior in the natural environment. Traditional full-call methods tend to overestimate the speed of policy response and the breadth of technology diffusion because they implicitly assume that everyone is constantly paying attention and ready to act; while the DDI mechanism, by introducing "decision inertia" and "information dependence," makes the simulation process closer to the complex picture of the coexistence of "silent majority" and "active adopters" in reality, thus providing a more realistic and explanatory sandbox environment for policy evaluation.

[0058] Specifically, the next decision time for each household intelligent agent i Due to its initial decision delay Decision interval with subsequent dynamic adjustments Joint decision:

[0059] in, This is the simulation start time. The time of the last decision.

[0060] (1) Initial delay: Initial decision delay This reflects the time required for a household to go from the start of the simulation to actively considering photovoltaic (PV) installation for the first time. According to the diffusion of innovation theory (Rogers, 2003) and the theory of planned behavior (Ajzen, 1991), the strength of an individual's attitude towards environmentally friendly technologies is a key antecedent driving their early information search and consideration behavior. Numerous empirical studies confirm that residents with stronger environmental awareness are more likely to pay attention to and evaluate the feasibility of distributed PV earlier, and early PV adopters are more driven by non-economic factors (such as environmental awareness).

[0061] Therefore, environmental awareness As a modulating variable, the following initial delay function is constructed:

[0062] in This represents the base random delay (in months), reflecting the uncertainty of information penetration in the early stages of the market. This serves as the median benchmark for environmental awareness. This is a sensitivity coefficient that controls the intensity of environmental awareness's influence on delay. This setting ensures that environmentally conscious families... The average delay is shorter, but individual heterogeneity is preserved, avoiding deterministic assumptions.

[0063] (2) Dynamic Decision Interval (DDI) To simulate the discontinuous nature of farmers' decisions regarding photovoltaic installation in reality, this embodiment constructs a Dynamic Decision Interval (DDI) mechanism. Decision Interval It is a weighted composite of multiple factors, including individual economic attributes, risk preferences, social structure, and information access capabilities.

[0064]

[0065] The relevant parameters and variables are defined as follows: , representing the baseline decision frequency (unit: month); The income level weighting coefficient is set to -1, 0, and 2 for high-, middle-, and low-income groups, respectively, to reflect the impact of economic redundancy on decision sensitivity. This represents the risk aversion level of the agent, with the corresponding weighting coefficient set as follows: P(propᵢ) represents the weight of the property constraint environment, specifically set as follows: -2 when the property supports it, 0 when there is no property such as in urban villages, and +4 when the property opposes it. The weighting coefficients represent the diversity of information channels accessed by an agent. This demonstrates how information completion can shorten the decision-making cycle.

[0066] After the agent makes a rejection decision, the system uses natural language processing technology to analyze the reasons for rejection. Heuristic fine-tuning of the decision interval based on semantics:

[0067] in, This is a predefined semantic adjustment function. If the rejection reason points to "high cost or price," then positive correction is performed. To simulate risk-averse waiting behavior; if the reason is attributed to "wait-and-see or insufficient information", a negative correction is made to reflect the agent's information-seeking motivation.

[0068] The generation of the structured rejection reasons does not rely on the LLM to directly output a preset format. Instead, the LLM first generates natural language reasons, and then maps them to a limited number of predefined rejection categories (such as 'cost factors', 'insufficient information', and 'risk aversion') through a rule-based keyword matching mechanism, and dynamically adjusts the decision interval according to the category attributes.

[0069] The semantic adjustment function The value is assigned based on the parsed rejection reason c, and the specific mapping and calibration rules include: Cost-sensitive mapping: If reason c contains keywords such as 'price', 'cost', 'expensive', or 'fee', it is identified as a 'cost factor'. The system calculates the delay step size based on the agent's decision patience value P (each agent's patience value is randomly given at initialization), using the formula δ = 6 + P / 2, and adjusts the decision interval upwards to simulate long-term wait-and-see attitude caused by high cost pressure.

[0070] Cautious wait-and-see mapping: If reason c contains keywords such as 'wait', 'wait', or 'let's see', it is identified as 'decision inertia'. The system calculates the shortening step size δ = max(2, 6 - P / 2) in reverse based on the patience value to simulate the transformation of wait-and-see sentiment into decision-making behavior over time.

[0071] Information-driven mapping: If reason c contains keywords such as 'information', 'understanding', 'consultation', or 'research', it is identified as 'insufficient information'. The system will shorten the decision interval by a fixed 2 months (δ=-2) to simulate the decision acceleration brought about by the agent actively searching for information.

[0072] Risk-averse mapping: If reason c contains keywords such as 'risk', 'worry', 'anxiety', or 'uncertainty', it is identified as 'risk aversion'. The system calculates the delay step δ = Ra / 2 based on the agent's own risk aversion coefficient Ra, thereby reconstructing the different responses of heterogeneous individuals to uncertainty at the micro level.

[0073] To prevent the feedback loop from getting out of control, the fine-tuned decision interval is limited by a preset lower limit Tmin and an upper limit Tmax, ensuring that the agent does not fall into extreme states of never making a decision or making decisions at high frequency.

[0074] (3) Generation of Potential Decision-Makers: To simulate the real dynamics of the distributed photovoltaic market in an urban context, this embodiment introduces a dynamic potential decision-making household generation function N(t) to simulate the number of new households entering the decision-making process each month. This setting is based on Rogers' technology diffusion curve, which posits that the social penetration of new technologies is not a static process, but rather exhibits a non-linear S-curve growth as the reach of information expands. In the early stages of the simulation, limited by low market awareness, the number of new decision-making households is relatively small each month; under the influence of policy incentives or neighborhood effects, households that were originally in a wait-and-see state begin to enter the decision-making process in a concentrated manner, forming a non-linear growth in the simulated base. This dynamic access mechanism not only maps the physical growth of a city's housing stock and population size, but also effectively solves the simulation distortion problem caused by the 'sample aging' problem (i.e., the remaining samples become increasingly difficult to convert as the simulation progresses) by continuously introducing new households with heterogeneity, thereby accurately assessing the market response characteristics under different policy interventions. Its mathematical expression is:

[0075]

[0076] Where K represents the monthly upper limit of the number of new decision-makers constrained by the city's housing stock. To align with the midpoint of the policy release cycle, k is a growth coefficient characterizing the information diffusion rate, while This represents random noise simulating disturbances in the real environment.

[0077] In summary, the DDI mechanism, by integrating individual traits, social structural constraints, cognitive feedback, and exogenous shocks, effectively overcomes the problems of fixed rule loops in traditional multi-agent systems (ABM) or the expensive, indiscriminate periodic calls of Large Language Models (LLM) agents. This mechanism can accurately characterize behavioral inertia and timing selection in low-frequency technology diffusion, providing a more biologically meaningful simulation paradigm for high-perception decision-making scenarios such as photovoltaic installation.

[0078] This invention operates in a simulation environment composed of multiple intelligent computing units. Each intelligent computing unit (i.e., an agent) is configured as a decision-making entity with a preset attribute profile, and each unit contains internal state parameters (such as economic level, cognitive tendency, etc.). In this scheme, the intelligent computing unit acts as the execution carrier of the DDI mechanism. Its core task is to receive environmental disturbance signals and, according to the dynamic decision interval algorithm described in this invention, autonomously determine the timing of the inference action. The unit structure here is a well-known or common large language model-driven structure in the art. The innovation of this invention lies in the dynamic adjustment of the decision frequency of this unit. When the dynamic decision interval (DDI) described in this invention reaches the trigger threshold, the intelligent computing unit is activated and enters the 'explicit reasoning state'. In this state, the unit uses a multi-step chain-of-thought logic to parse the current environmental information. It should be noted that the chain-of-thought process is only a downstream execution link after the DDI mechanism is triggered, used to output the final decision result. This invention focuses on how to filter out invalid reasoning timing points through the DDI mechanism, thereby avoiding redundant chain-of-thought by the agent when there are no changes in key information. A key feature of this invention is that the computational depth of the dynamic decision interval integrates historical feedback signals from the decision-making entity. Specifically, when the intelligent computing unit generates a 'rejection' signal in the previous decision cycle, the semantic reflection content carried by this signal (i.e., Reflexion reflection information) is converted into numerical cognitive inertia parameters. The cognitive inertia parameter is determined by the semantic adjustment function. δ ( RThe correction value generated represents the numerical impact of historical reflection information on the agent's future decision-making timing. This parameter, as a crucial input to the core formula of this algorithm, directly influences the next decision interval. The length of the decision window. In this way, the DDI mechanism can identify decision stagnation caused by 'cognitive rigidity' and dynamically lengthen or shorten the decision window accordingly, thereby achieving accurate simulation of the 'low-frequency decision' characteristics in the real world.

[0079] Experimental verification To verify the effectiveness of this invention in a photovoltaic adoption scenario in a certain city, the simulation period was set from 2020 to 2025. Figure 3 The graph shows the monthly evolution curves of newly added potential decision-making households generated by the logistic model: the dark gray scatter dots represent simulated values ​​with a 5% random disturbance, aiming to reproduce the uncertainty of the real market; the black dashed line reflects the overall growth trend of the potential sample pool. In terms of parameter settings, the saturation capacity K=45 is set based on the growth level of housing stock, and the growth midpoint t0=2023.5 precisely matches the release node of a city's photovoltaic incentive policy. This dynamic access mechanism not only maps the physical growth of urban space but also effectively solves the simulation distortion caused by "sample aging" in the later stages of the simulation by continuously introducing heterogeneous samples, ensuring the model's authenticity and sensitivity in policy extrapolation.

[0080] To ensure the predictive accuracy and generalization ability of the model framework, historical installed capacity data from January 2020 to December 2023 was used as the training and calibration set to fine-tune the initial attribute distribution of the agent (such as environmental awareness and risk preference) and the hyperparameters in the DDI mechanism. Subsequently, data from January 2024 to December 2025 was used as an independent test set. In this stage, the model underwent autonomous simulation without masking real-world results. By comparing the simulated installed capacity curve with actual statistical data, the mean total error and dynamic evolution correlation were calculated to verify the model's robustness in the face of policy fluctuations and social network evolution.

[0081] The open-source large language model GPT-OSS-120B was used as the agent inference engine, with a uniform temperature coefficient of 0.7 during invocation. A deterministic seed sequence was used to control five repeated simulation experiments. Specifically, all random seeds started from a base value of 42 and increased sequentially in fixed increments of 100.

[0082] To ensure the reliability and generalizability of the simulation results, the simulation was repeated five times under the same parameter settings to generate average fitted data. The final average fitted value is as follows: Figure 4As shown, the simulated installation curves exhibit a high degree of consistency with the actual trends. The model accurately captured the period of steady growth driven by subsidy policies from 2021 to 2022, as well as the explosive growth driven by policy expectations in the first half of 2025. This demonstrates that policy responses captured through a dynamic decision-making mechanism can effectively reconstruct the complex process of socio-technology diffusion.

[0083] To quantify the fitting accuracy, the Annual Total Volume Error (ATVE) is used as the evaluation metric. Since the installation of distributed photovoltaic systems fluctuates in the short term (e.g., monthly) due to random factors such as weather and construction cycles, the key to evaluating the model's effectiveness lies in its ability to capture the overall annual installation scale.

[0084] First, calculate the relative error between the simulated total and the actual total for each year:

[0085] Subsequently, by averaging the errors for each year within the observation period (2020-2025), the final comprehensive evaluation index was obtained:

[0086] Where y represents the year, m represents the month, and Y is the total number of years. y and y' represent the simulated and actual values, respectively. The calculation results are shown in Table 1. The final five-year average error is only 5.16%, verifying the model's excellent performance in long-term aggregate prediction.

[0087] To verify the effectiveness of each component in the framework proposed in this invention, five sets of comparative experiments were designed. By conducting a stripping experiment on the Dynamic Decision Interval (DDI), Semantic Fine-tuning (SFT), and Reflective Memory (RM), the contribution of each module to the simulation accuracy was analyzed.

[0088] Table 1: Experimental Groups and Their Meanings

[0089] The detailed configurations of each experimental group are shown in Table 1. Among them, the "fixed step size existing technology" (referencing Chetty's 2025 simulation framework based on residential electricity consumption) is used as the baseline scheme. It adopts a residential energy simulation framework based on large language model (LLM) agents, but uses fixed periodic updates (such as monthly updates) in decision scheduling.

[0090] Table 2: Summary of Quantitative Results of Ablation Experiments (2020-2025)

[0091] Table 2 summarizes the quantitative indicators of each group of experiments from 2020 to 2025, and records the annual value and five-year average of the total capacity error (ATVE).

[0092] As shown in Table 2 and Figure 5 As shown, both the fixed-step comparison scheme (including the reflection module but excluding DDI) and the fixed-step basic agent (excluding DDI and the reflection module) exhibit significant prediction biases, with average annual total error (ATVE) reaching 81.65% and 107.21%, respectively. Comparative analysis reveals that the existing comparison scheme, under its fixed-time-step forced wake-up mechanism, compels the computing unit to execute decision-making logic monthly for high-value photovoltaic investments. During periods of policy stability, frequent rejection feedback is continuously reinforced by the reflection module, leading to "cognitive rigidity" in the computing unit. This makes the agent extremely insensitive to real market opportunities in the later stages, causing decision-making activity to rapidly drop to zero after the initial phase. This explains why the model without the reflection mechanism slightly outperforms the existing comparison scheme with the reflection mechanism on certain indicators. Therefore, this type of comparison scheme is more suitable for high-frequency consumption simulations (such as daily electricity generation) rather than low-frequency investment simulations.

[0093] In summary, this invention proposes a dynamic simulation method for low-frequency decision-making based on a large language model. This method aims to address the problems of cognitive rigidity, decision frequency distortion, and wasted computational resources caused by the fixed-step triggering mechanism in existing large-model simulation frameworks when handling "high-cost, low-frequency" decisions such as residential photovoltaic adoption. The core of its technical solution lies in constructing a dynamic decision interval mechanism: achieving agent heterogeneity through initial delay allocation, dynamically calculating decision trigger points by combining external policy disturbances and internal cognitive evolution, and utilizing a natural language-based reflection-based storage module to enable the agent to self-correct the decision interval based on historical rejection reasons (such as cost or insufficient information), thereby achieving high-fidelity simulation of decision inertia.

[0094] This invention addresses the technical bottlenecks of existing social behavior simulation technologies based on large language models when dealing with "low-frequency, high-cost, long-cycle" decision-making scenarios (such as distributed photovoltaic installation, home purchase, and large equipment upgrades), and proposes a systematic solution. Existing technologies mainly suffer from the following problems: First, misaligned decision-making frequency and cognitive rigidity. The fixed-step full-activation mechanism forces the agent to make decisions at every time point, leading to frequent "rejection" memories and pathological self-reinforcement in low-frequency scenarios, preventing the model from responding correctly when real opportunities arise later. Second, dynamic distortion of simulation time. Traditional models ignore the "decision inertia" and "observation period" of residents before major investments, making it difficult to depict the nonlinear process of decision windows triggered by external information disturbances, resulting in significant deviations in macroscopic diffusion rate prediction. Third, waste of computational resources. In large-scale urban simulations, the full-cycle periodic invocation of large language models incurs high computational costs, most of which are invalid and repeated invocations. Compared with existing technologies, the significant technical advantages of this invention are reflected in the following aspects:

[0095] 1. This invention breaks through the unrealistic assumption of "all-person decision-making" in traditional simulations, achieving precise screening and mapping of the decision-making population. Traditional simulation methods typically assume that all individuals are in a decision-making state at every time step, which is seriously inconsistent with the "low-frequency, non-active attention" characteristics of decision-making in reality. This invention, through a dynamic decision-making interval mechanism, activates the explicit reasoning of the corresponding agent only when it perceives external environmental disturbances or internal cognition reaches a trigger point. This ensures that the active distribution of the agent group accurately corresponds in time and space to the real population "generating decision-making intentions." This "on-demand activation" logic allows the model to more realistically reproduce the dynamic process of alternating "latent period" and "explosive period" in the diffusion of social technology, rather than a mechanical linear growth.

[0096] 2. Significantly reduces the consumption of computing resources in large-scale social simulation. Since it no longer requires all agents to call large language models for reasoning at every time step, but instead accurately identifies active individuals through dynamic decision intervals, redundant calling overhead (API requests or GPU computing power overhead) is greatly reduced in simulations of the same population size, making it possible to perform refined behavioral simulations at the city level or even on the scale of tens of millions of people with limited resources.

[0097] 3. Significantly enhanced cognitive consistency of agents in long-term simulations. By combining dynamic decision intervals with a reflection-based storage module, the problem of agent state jumps or memory loss during simulations lasting several years is effectively solved. The dynamic interval provides reasonable evolution time for semantic correction, ensuring that the agent's psychological evolution trajectory is highly consistent with the prudent decision-making characteristics of real humans when the decision span is large.

[0098] This invention also provides a storage medium for storing a computer program, which, when executed, performs at least the methods described above.

[0099] This invention also provides a control device, including a processor and a storage medium for storing a computer program; wherein the processor executes the computer program by performing at least the method described above.

[0100] This invention also provides a processor that executes a computer program, at least performing the methods described above.

[0101] The storage medium can be implemented by any type of non-volatile storage device, or a combination thereof. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc or CD-ROM; magnetic surface memory can be disk storage or magnetic tape storage. The storage media described in the embodiments of this invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0102] In the several embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0103] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0104] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0105] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0106] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0107] The methods disclosed in the several method embodiments provided by this invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0108] The features disclosed in the several product embodiments provided by this invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0109] The features disclosed in the several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0110] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various equivalent substitutions or obvious modifications can be made without departing from the concept of the present invention, and all such modifications, achieving the same performance or application, should be considered within the scope of protection of the present invention.

Claims

1. A dynamic simulation method for low-frequency decision-making based on a large language model, used to simulate the diffusion trend of low-frequency, high-cost decision-making behavior, characterized in that, Includes the following steps: S1. Construct a heterogeneous group of intelligent agents and assign an initial decision delay to each agent to simulate the difference in the time of the first consideration of the target decision by the individual at the beginning of the simulation. S2. Monitor external environmental disturbances and internal cognitive states in real time, dynamically calculate the decision interval of each agent, and trigger explicit reasoning of the agent based on the decision interval or external disturbances, so that the agent only enters the decision state when the decision interval is zero or when it is subjected to a major external disturbance. S3. For the triggered agent, call the large language model to perform decision reasoning. If the decision result is rejection, generate a structured rejection reason and store it in the memory module. In subsequent decision cycles, the agent retrieves and reflects on historical rejection reasons to ensure the consistency of long-term decision logic. S4. Summarize the decision results of all agents and generate simulation output to predict the diffusion trend of low-frequency, high-cost behaviors.

2. The low-frequency decision dynamic simulation method based on a large language model as described in claim 1, characterized in that, In step S1, the allocation of the initial decision delay is based on the innovation diffusion theory combined with individual characteristics: the basic delay is generated by random distribution and adjusted according to the difference of each agent's environmental awareness relative to the preset benchmark level. The higher the environmental awareness, the shorter the initial delay, so as to simulate the heterogeneity of individuals' attention time to new technologies.

3. The low-frequency decision dynamic simulation method based on a large language model as described in claim 1, characterized in that, In step S2, the dynamic calculation of the decision interval for each agent includes: based on a preset baseline decision interval, and taking into account the agent's economic attributes, risk preferences, social structural constraints, and information acquisition capabilities, the decision interval value is obtained by weighted summation; wherein, factors such as high income, low risk aversion, property support, and multiple information channels are assigned weights to shorten the decision interval, and vice versa.

4. The low-frequency decision dynamic simulation method based on a large language model as described in claim 1, characterized in that, In step S2, the explicit reasoning of the triggering agent includes: automatic activation when the decision interval counter returns to zero, or immediate activation when a major disturbance event occurs in the external environment, so that the agent enters the decision state.

5. The low-frequency decision dynamic simulation method based on a large language model as described in claim 1, characterized in that, In step S3, generating structured rejection reasons and storing them in the memory module includes: the activated agent calls the large language model to perform chained reasoning; if a rejection decision is generated, the model is guided to output a formatted rejection reason and store it in long-term memory; in subsequent decision cycles, the agent retrieves historical rejection reasons from the memory module and uses them as context input to the large language model to reflect on and maintain decision consistency.

6. The low-frequency decision dynamic simulation method based on a large language model as described in claim 1, characterized in that, In step S3, after reflecting on the historical reasons for rejection, the reflected information is converted into a numerical cognitive inertia parameter. This parameter is used to adjust the length of the next decision interval, so that the decision frequency of the agent is affected by historical cognitive feedback.

7. The low-frequency decision dynamic simulation method based on a large language model as described in claim 1, characterized in that, In step S3, after generating the structured rejection reasons, the next decision interval is further fine-tuned based on the semantic content of the rejection reasons: if the rejection reason points to high cost, the decision interval is extended to simulate risk-averse waiting behavior; if the rejection reason points to insufficient information, the decision interval is shortened to simulate information search motivation.

8. The low-frequency decision dynamic simulation method based on a large language model as described in claim 1, characterized in that, In step S4, the simulation output includes generating a technology adoption diffusion curve. By comparing the curve changes under different policy interventions, the incentive effect of policies on low-frequency decision-making behavior is evaluated.

9. The low-frequency decision dynamic simulation method based on a large language model as described in claim 1, characterized in that, After step S1 or during the simulation, a step of dynamically introducing new agents is also included: according to the technology diffusion curve model and combined with random perturbations, the number of new agents entering the decision-making process is generated month by month to simulate the continuous emergence of potential adopters in the real market and avoid simulation distortion caused by sample aging.

10. The low-frequency decision dynamic simulation method based on a large language model as described in claim 1, characterized in that, The method is applied to the accurate prediction and policy simulation scenarios of residential photovoltaic adoption behavior to simulate the decision-making process of residents when facing electricity price fluctuations, subsidy policy adjustments, and neighborhood installation effects; or, the method is applied to the home purchase decision simulation scenario to simulate the timing and hesitation behavior of families when facing changes in credit policies, regional development plans, and market expectation disturbances; or, the method is applied to the education investment decision simulation scenario to simulate the decision delay and final choice behavior of families when considering their children's long-term education investment, influenced by policy guidance, social comparison, and information access capabilities.

Images