KR102153540B1 - Method and apparatus for micro simulation parameter calibration using machine learning in agent based simulation - Google Patents

Abstract

A method for calibrating micro-simulation parameters using machine learning in an agent-based simulation performed by a computer is provided. The method of the present disclosure includes the steps of: (a) generating validation data by pre-processing moving target data, which is target data to be estimated through simulation, including characteristic data of a plurality of agents, and converting it into a predetermined data format; (b) Agent-Based Modeling and Simulation (ABMS) is performed using scenario data for reflecting the initial parameter set and external variables for the characteristics of the plurality of agents in the simulation as ABM input data. Obtaining agent micro data of each of the plurality of agents by performing, and obtaining agent micro data by aggregating the agent micro data; (c) performing micro-clustering analysis on micro-data of each of the plurality of agents to generate at least one clustering and performing cluster analysis; (d) analyzing an error by comparing the validation data and the agent macroscopic data; (e) 　, when the error exceeds a predetermined reference value, (e-1) 　 calibrate the parameters for each of the at least one cluster to reduce the error, and set a calibration parameter set for each of the at least one cluster Setting up; (e-2) Each of the plurality of agents by setting the calibration initial parameter set for the characteristics of the plurality of agents and the scenario data as ABM input data, and performing agent-based modeling and simulation on the ABM input data. Obtaining agent micro data of the agent and collecting the agent micro data to obtain agent macro data; (e-3) performing steps (d) and (e); (f) when the error is less than or equal to a predetermined reference value, determining the corrected parameter as a final parameter.

Description

Method and apparatus for micro simulation parameter calibration using machine learning in agent based simulation}

The present disclosure relates to an agent-based simulation method, and more particularly, to a micro-simulation parameter calibration method using machine learning in agent-based simulation.

It is a chronic problem of agent-based simulation that the result of agent-based simulation is different from the observation result of the system to be modeled. In order for a simulation model with a generative feature to solve this problem through parameter correction to be used universally, the reliability of the model must be guaranteed. In order to increase the reliability of the model, it is possible to calibrate model parameters based on the generated simulation result. Machine learning, that is, correction of simulation parameters through data-based statistical models, is key to increasing the reliability of the simulation model.

Previous work on parameter calibration was generally a methodology applicable only to specific simulation scenarios. A more general simulation calibration methodology is a method using a genetic algorithm. However, there is a limitation in that a method using a genetic algorithm can be applied only when a social network is modeled in a simulation model in order to use the genetic algorithm. In addition, the method using the genetic algorithm learns the parameters that determine the strategy based on the utility of the agent, but it has the disadvantage that it is very difficult to model the utility of the agent.

As a more general simulation model calibration methodology, a dynamic parameter calibration methodology using machine learning ("Data-Driven Automatic Calibration for Validation of Agent-Based Social Simulations") has been introduced. This is a methodology for learning dynamic parameters in the direction of directly increasing the fit of the simulation, and has the advantage that it can be applied to all simulation models. For convenience, when the corrected parameter in the methodology is called a macro parameter, the role of the macro parameter means a general parameter that can be applied to all agents in common. After detecting a regime using a hidden Markov model for dynamic macro parameters, the macro parameters are estimated for each regime. In other words, it is a dynamic macroscopic parameter calibration methodology that changes over time. However, the factors that determine the agent's strategy include common factors and other factors of each agent. Therefore, a correction methodology is required for micro-parameters that determine factors of each agent.

[1] B. Park and H. Qi, “Development and evaluation of a procedure for the calibration of simulation models,” Transportation Research Record: Journal of the Transportation Research Board (TRB), 2005. [2] T. Toledo, M. Ben-Akiva, D. Darda, M. Jha, and H. Koutsopoulos, “Calibration of microscopic traffic simulation models with aggregate data,” Transportation Research Record: Journal of the Transportation Research Board (TRB ), 2004. [3] J. Hourdakis, P. Michalopoulos, and J. Kottommannil, “Practical procedure for calibrating microscopic traffic simulation models,” Transportation Research Record: Journal of the Transportation Research Board (TRB), 2003. [4] V. Nannen and A. E. Eiben, “A method for parameter calibration and relevance estimation in evolutionary algorithms,” In 8th annual conference on Genetic and evolutionary computation (ACM), 2006. [5] Il-Chul Moon, Dongjun Kim, Tae-Sub Yun, Jang Won Bae, Dong-oh Kang, and Euihyun Paik, “Data-Driven Automatic Calibration for Validation of Agent-Based Social Simulations,” In IEEE International Conference on Systems, Man, and Cybernetics (SMC), 2018

There is a need for a method to increase the reliability of the simulation model.

According to an aspect of the present disclosure, a method of calibrating micro-simulation parameters using machine learning in agent-based simulation, performed by a computer, is provided. The method of the present disclosure includes the steps of: (a) generating validation data by pre-processing moving target data, which is target data to be estimated through simulation, including characteristic data of a plurality of agents, and converting it into a predetermined data format; (b) Agent-Based Modeling and Simulation (ABMS) is performed using scenario data for reflecting the initial parameter set and external variables for the characteristics of the plurality of agents in the simulation as ABM input data. Obtaining agent micro data of each of the plurality of agents by performing, and obtaining agent micro data by aggregating the agent micro data; (c) performing micro-clustering analysis on micro-data of each of the plurality of agents to generate at least one clustering and performing cluster analysis; (d) analyzing an error by comparing the validation data and the agent macroscopic data; (e) 　, when the error exceeds a predetermined reference value, (e-1) 　 calibrate the parameters for each of the at least one cluster to reduce the error, and set a calibration parameter set for each of the at least one cluster Setting up; (e-2) Each of the plurality of agents by setting the calibration initial parameter set for the characteristics of the plurality of agents and the scenario data as ABM input data, and performing agent-based modeling and simulation on the ABM input data. Obtaining agent micro data of the agent and collecting the agent micro data to obtain agent macro data; (e-3) performing steps (d) and (e); (f) When the error is less than or equal to a predetermined reference value, determining the corrected parameter as a final parameter may be included.

According to another feature of the present disclosure, a method of calibrating micro-simulation parameters using machine learning in agent-based simulation, performed by a computer, is provided. The method of the present disclosure includes the steps of: (a) generating validation data by pre-processing moving target data, which is target data to be estimated through simulation, including characteristic data of a plurality of agents, and converting it into a predetermined data format; (b) Agent-Based Modeling and Simulation (ABMS) is performed using scenario data for reflecting the initial parameter set and external variables for the characteristics of the plurality of agents in the simulation as ABM input data. Obtaining agent micro data of each of the plurality of agents by performing, and obtaining agent micro data by aggregating the agent micro data; (c) performing micro-clustering analysis on micro-data of each of the plurality of agents to generate at least one clustering and performing cluster analysis; (d) analyzing an error by comparing the validation data and the agent macroscopic data; And (e) when the error exceeds a predetermined reference value, estimating a micro parameter.

In one embodiment, the cluster analysis of the above-described method is performed by a method of performing cluster analysis through latent expressions, such as a variant autoencoder (VAE) and a dirichlet process mixture model (DPMM). Can be applied.

According to another feature of the present disclosure, a computer-readable recording medium including one or more instructions, wherein the one or more instructions, when executed for a computer, cause the computer to perform any one of the above-described methods. A computer-readable recording medium may be provided.

According to another feature of the present disclosure, an apparatus for calibrating micro-simulation parameters using machine learning in agent-based simulation may be provided. The apparatus of the present disclosure includes characteristic data of a plurality of agents, and a validation data generation module configured to generate validation data by preprocessing the moving target data, which is target data to be estimated through simulation, and converting it into a predetermined data format. ; Agent-Based Modeling and Simulation (ABMS) is performed using scenario data for reflecting the initial parameter set and external variables for the characteristics of the plurality of agents in the simulation as ABM input data, A modeling module configured to obtain agent micro data of each of a plurality of agents, and to obtain agent macro data by aggregating the agent micro data; A cluster analysis module configured to generate at least one cluster by performing micro cluster analysis on the micro data of each of the plurality of agents and perform cluster analysis; An error analysis module configured to analyze an error by comparing the validation data and the agent macroscopic data; And a micro parameter estimation module configured to calibrate the parameter for each of the at least one cluster to reduce the error and set a calibration parameter set for each of the at least one cluster.

According to the present disclosure, it is possible to increase the reliability of a simulation model by selecting a parameter through micro-simulation parameter calibration rather than using a parameter using a heuristic technique.

According to the present disclosure, a plurality of statistical models, i.e., a probabilistic self-coding model and a Dirichlet process mixture model for cluster analysis, and Bayesian optimization to obtain a micro parameter that derives the maximum degree of matching, are used to produce a high degree of matching. You can find the parameters.

1 is a diagram illustrating a method of calibrating micro-simulation parameters using machine learning according to an embodiment of the present disclosure.
2 is a flowchart illustrating a method of calibrating micro-simulation parameters using machine learning according to an embodiment of the present disclosure.
3 is a diagram illustrating a result of analyzing a latent expression for each cluster using a probabilistic self-encoder according to an embodiment of the present disclosure.
FIG. 4 is a graph showing results of simulation matching using a method of calibrating micro-simulation parameters using machine learning according to an embodiment of the present disclosure.
FIG. 5 is a graph of a metropolitan area apartment trading volume obtained by using a calibrated parameter based on a method of calibrating micro-simulation parameters using machine learning according to an embodiment of the present disclosure.
6 is a graph of a metropolitan area apartment trading volume calculated using an existing ad-hoc parameter.

Advantages and features of the present disclosure, and a method of achieving them will become apparent with reference to the embodiments described later in detail together with the accompanying drawings. However, the present disclosure is not limited to the embodiments disclosed below, but will be implemented in a variety of different forms, and the present embodiments only make the disclosure of the present disclosure complete, and those skilled in the art to which the present disclosure pertains. It is provided to inform the person of the scope of the invention completely, and the present disclosure is only defined by the scope of the claims.

The terms used in this specification are only used to describe specific embodiments and are not intended to limit the present disclosure. For example, a constituent element expressed in the singular should be understood as a concept including a plurality of constituent elements, unless the context clearly means only the singular. In addition, in the specification of the present disclosure, terms such as'include' or'have' are only intended to designate the existence of features, numbers, steps, actions, components, parts, or a combination thereof described in the specification. The use of the term does not exclude the possibility of the presence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof. In addition, in the embodiments described in the present specification, the'module' or'unit' may mean a functional part that performs at least one function or operation.

In addition, unless otherwise defined, all terms used herein including technical or scientific terms have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs. Terms as defined in a commonly used dictionary should be interpreted as having a meaning consistent with the meaning in the context of the related technology, and interpreted as an ideal or excessively formal meaning unless explicitly defined in the specification of the present disclosure. It doesn't work.

Hereinafter, embodiments of the present disclosure will be described in more detail with reference to the accompanying drawings. However, in the following description, when there is a possibility that the subject matter of the present disclosure may be unnecessarily obscure, detailed descriptions of widely known functions or configurations will be omitted.

1 is a diagram illustrating a method of calibrating micro-simulation parameters using machine learning according to an embodiment of the present disclosure.

In one embodiment of the present disclosure, the micro parameter means a parameter different for each agent that determines the strategy of each agent. However, correcting the parameter values of all agents may correspond to machine learning overfitting. Accordingly, in an embodiment of the present disclosure, the micro parameter is defined as a parameter of an agent cluster, and the micro parameter correction is defined as correcting a parameter allocated for each cluster through agent cluster analysis.

In an embodiment of the present disclosure, the moving object data may mean target data to be estimated through simulation.

In an embodiment of the present disclosure, the scenario data is data for reflecting external variables in the simulation. For example, in a housing simulation model that predicts how the housing price and transaction volume will move, the scenario data is housing policy variables such as LTV and DTI. Can be

Pre-processing of moving target data (110)

The moving object data may be converted into a data format that enables error analysis by preprocessing real data as target data to be estimated through simulation. In an embodiment of the present disclosure, in order to analyze a specific phenomenon through simulation, data of the phenomenon may be preprocessed and converted into the same format as the simulation result. In other words, it is possible to preprocess the real data generated based on the given parameters to be compared with the simulation results. This makes it possible to directly compare the real data and the simulation data, and through this, it is possible to find a simulation parameter that better simulates the real data.

As shown in FIG. 1, a graph 120 of real world data (validation data) may be obtained by pre-processing the moving object data.

Agent-based modeling and simulation (130)

Agent-Based Modeling and Simulation (ABMS) is capable of realizing detailed behaviors and interactions of agents that cannot be expressed by simulations based on modeling at the existing system level. It is used as a methodology for modeling in various fields such as diffusion prediction. In an embodiment of the present disclosure, the agent-based model (ABM) in the ABMS 130 shown in FIG. 1 refers to modeling on the premise of a multi agent system, and the agent performs optimal behavior. It is premised.

In one embodiment of the present disclosure, the agent micro data 140 is a simulation of the ABMS 130 as micro data for each agent.

In one embodiment of the present disclosure, the agent macroscopic data graph 150 is a graph in which agent microscopic data 140, which is a result value simulated by the ABMS 130, is aggregated and expressed as agent macroscopic data. The X axis of the graph 150 indicates time, and the Y axis indicates a simulation result.

Simulation error analysis (160)

Next, in an embodiment of the present disclosure, a statistically meaningful error may be obtained by comparing the graph 120 and the graph 150. In an embodiment of the present disclosure, a method of obtaining an error may use a root mean square error (RMSE) or a mean absolute percentage error (MAPE), but the present disclosure is not limited thereto. .

Micro cluster analysis (170)

In an embodiment of the present disclosure, after learning a hidden representation of an agent through a variant autoencoder (VAE), cluster analysis may be performed using the latent representation as a new input value. In an embodiment of the present disclosure, cluster analysis may be performed using a Dirichlet process mixture model (dpmm), but the present disclosure is not limited thereto.

In an embodiment of the present disclosure, in the case of a housing model with 10,000 agents, 250-dimensional data exists for each agent, and a 10-dimensional latent expression can be obtained for 10,000 agents. That is, 250-dimensional data (saving, income, and debt of the agent) can be compressed into 10 dimensions. Next, a cluster analysis is performed on 10000 10-dimensional data using a clustering algorithm, for example, a Dirichule process mixed model to determine which cluster each 10,000 agents belong to.

For example, an individual can buy or sell a house for several reasons. When an individual buys a house, he or she can buy a house based on the overall economic situation, personal circumstances, for example, because the house is indispensable, so he can buy a house, or he can buy a house in order to receive a small amount each month. In addition, when an individual says they are selling a house, they can see the overall economic situation and sell the house, or they can sell the house because they need a small amount of money. In other words, buying and selling a house may differ according to individual investment preferences. Although the criteria for judging certain behaviors for each individual may be different, the criteria for judging each individual can be clustered into micro parameters and analyzed.

In an embodiment of the present disclosure, if the cluster analysis is performed once for the first time, it may not be performed again during calibration iteration.

Micro parameter estimation (180)

Next, in an embodiment of the present disclosure, parameters may be assigned to agents belonging to each cluster. If they belong to the same cluster, the same parameters are assigned. The assigned parameter becomes the input parameter of the next simulation.

In an embodiment of the present disclosure, the micro-parameter estimation 180 may use a Gaussian process using Bayesian optimization, but the present invention is not limited thereto.

In one embodiment of the present disclosure, the calibration iteration may be repeated until the simulation result graph 150 finds a set of microscopic parameters that best match the real data 120. In one embodiment of the present disclosure, when the simulation error analysis 160 is less than or equal to a predetermined reference value, the corrected micro parameter set may be regarded as the final parameter.

2 is a flowchart illustrating a method of calibrating micro-simulation parameters using machine learning according to an embodiment of the present disclosure.

를 설정한다. 여기서, k는 교정하고자 하는 미시 파라미터의 인덱스(index)를 의미하고, 0는 금번 파라미터 세트가 초기 파라미터 세트임을 의미한다. 초기 파라미터 세트로 시뮬레이션을 수행하였을 때, 시뮬레이션의 에이전트 별 미시적 특성은 시간에 따른 결과값

을 가진다. 여기서, t는 시간을, att는 특성을 의미한다. First, the initial parameter set in step S210

Is set. Here, k means the index of the micro parameter to be corrected, and 0 means that this parameter set is an initial parameter set. When the simulation is performed with the initial parameter set, the microscopic characteristics of each agent in the simulation are the result values over time.

Have. Here, t stands for time and att stands for characteristics.

로 시뮬레이션을 반복 실험할 수 있다. 본 개시의 일 실시예에서, 새로운 파리미터 세트로 시뮬레이션을 반복하기 위해 파라미터 세트를 에이전트 별 미시적 특성의 시간에 따른 결과값

으로 치환할 수 있다. 그 후, 단계(S230)에서, 에이전트 별 미시적 특성의 시간에 따른 결과값

을 입력으로하여 차원 축소(dimension reduction) 기법을 통한 잠재표현

를 구할 수 있다. 본 개시의 일 실시예에서, 차원 축소 기법으로 기계 학습 방법론인 확률적 자기부호화기를 사용할 수 있으나 이에 한정되는 것은 아니다. Next, in step S220, a new parameter set

The simulation can be repeated. In an embodiment of the present disclosure, in order to repeat the simulation with a new parameter set, a parameter set is used as a result of the microscopic characteristics of each agent over time.

Can be substituted with Then, in step (S230), the result value according to the time of the microscopic characteristics of each agent

Latent expression through dimension reduction technique using as input

Can be obtained. In an embodiment of the present disclosure, a probabilistic self-encoder, which is a machine learning methodology, may be used as a dimensionality reduction technique, but is not limited thereto.

를 입력으로 하여 군집(cluster)을 구성할 수 있다. 군집 구성 후, 각 에이전트가 어떤 군집에 속하는지 알 수 있다. Next, in step S240, the latent expression obtained for each agent

A cluster can be formed by using as an input. After clustering, it is possible to know which cluster each agent belongs to.

In one embodiment of the present disclosure, the cluster configuration may use a Dirichlet process mixing model. The Dirichlet process mixed model is a nonparametric cluster analysis methodology. The Dirichlet process hybrid model determines which cluster the data will belong to, given the data, according to the Chinese restaurant rule. Here, according to the Chinese restaurant rule, the probability of belonging to the existing cluster when new data is input is proportional to the number of data belonging to the cluster, and the probability of forming a new cluster is proportional to the model variable γ. After generating the posterior distribution by multiplying the prior distribution generated by the Chinese restaurant rule by the likelihood of the new data, the cluster of data can be determined through sampling.

In another embodiment of the present disclosure, a Gaussian mixed model, which is a cluster analysis method with known parameters, may be used. In the Gaussian mixed model, when the simulation model changes, the number of agent clusters can also change. Therefore, an appropriate number of clusters should be determined for each simulation model.

As shown in FIG. 3 to be described later, it can be confirmed that agents located in the same place belong to the same cluster.

로 설정할 수 있다. 여기서 c는 군집을 의미하고, k는 교정하고자 하는 미시 파라미터의 인덱스를 의미하며, i는 파라미터 교정 반복 횟수를 의미한다. In step S250, a parameter set for each cluster is

Can be set to Here, c denotes a cluster, k denotes the index of a micro parameter to be corrected, and i denotes the number of repetitions of parameter correction.

로 설정할 수 있다. 즉, 단계(S250)에서 군집 별 파라미터 세트

를 입력으로 하여 시뮬레이션을 수행하면, 정합도 계산에 쓰이는 시뮬레이션 결과

를 구할 수 있다. 여기서, t는 시간을 의미한다. In step S260, the parameter set is simulated

Can be set to That is, in step S250, the parameter set for each cluster

If the simulation is performed with as input, the simulation result used for the matching degree calculation

Can be obtained. Here, t means time.

In step S270, Bayesian optimization may be performed by using the simulation matching degree. In an embodiment of the present disclosure, after calculating the degree of matching by comparing with validation data, Bayesian optimization is performed based on the degree of matching to estimate a set of calibration microparameters having a maximum matching degree. . Here, the verification data is data that enables error analysis by preprocessing actual data.

를 미시 파라미터로,

를 정합도로 가지는 데이터

가 있을 때, 데이터를 기반으로 가우시안 프로세스를 진행하여 함수 y=f(x)의 모양을 추정하고, 함수의 모양을 기반으로 최댓값이 나올 확률이 가장 높은 미시 파라미터 값을 추정할 수 있다. 추정한 미시 파라미터 값으로 시뮬레이션을 진행하여 새로운 정합도를 구한 후, 새로운 데이터

를 기존 데이터에 추가하여 위의 과정을 반복한다.In an embodiment of the present disclosure, the Bayesian optimization method is a methodology for finding a micro parameter value having a maximum degree of matching when there is a function that takes micro parameters as an input and outputs the degree of matching. Bayesian optimization method is currently

As a micro parameter,

Data with a degree of consistency

When is present, a Gaussian process is performed based on the data to estimate the shape of the function y=f(x), and the microparameter value with the highest probability of obtaining the maximum value can be estimated based on the shape of the function. After the simulation is performed with the estimated microparameter values, a new degree of matching is obtained, and the new data

Is added to the existing data and the above process is repeated.

Given the data, the microparameter value having the maximum degree of matching can be found by using an Expected Improvement acquisition function. For an arbitrary parameter x, the value of the Expected Improvement acquisition function (EI) is the sum of calculating the probability that f(x) is higher than the maximum matching degree estimated by the current Gaussian process, and weighting it as high as possible. . Bayesian optimization is to set the micro parameter x where the acquisition function is maximized as a new parameter.

[Equation 1]

In step S280, if the estimated new micro parameter set (corrected parameter set) is input and the above steps are repeatedly performed, the maximum matching degree can be finally obtained.

3 is a diagram illustrating a result of analyzing a latent expression for each cluster using a probabilistic self-encoder according to an embodiment of the present disclosure.

Here, the probabilistic self-encoder is a machine learning methodology that finds the latent representation of data. There is a drawback that the latent representation of data is generally uninterpretable, but research results have known that it is more effective to perform cluster analysis through the latent representation for large amounts of data with large dimensions.

Probabilistic self-encoder is a self-encoder that assumes that the prior distribution followed by the latent representation of the latent space is a standard Gaussian distribution. The result of the encoder part comes out as the mean and variance of the latent representation of the input x, and the latent representation of the input x is obtained by sampling from the Gaussian distribution with the mean and variance, which are the encoder results. A reparametrization trick can be used to solve the problem of inability to differentiate an encoder partial neural network due to sampling.

The probabilistic self-encoder can be learned using an evidence lower bound (ELBO), which is a lower limit of the loglikelihood of data, as a gradient descent method.

[Equation 2]

A latent expression of the learning result data comes out, and when the latent expression is expressed in two dimensions using a t-sne visualization method, it is shown as in FIG. 3. As shown in FIG. 3, it can be seen that data of the same cluster have similar latent expression values. FIG. 3 is a diagram illustrating a result of extracting a 10-dimensional latent expression of 240-dimensional data, and then compressing it into two dimensions through a t-sne method and visualizing it.

Simulation model

The agent-based simulation according to an embodiment of the present disclosure may be a housing simulation model. Here, the housing simulation model is a model for analyzing the housing market fluctuations according to policy changes. Through the housing simulation model, it is possible to analyze the price index and liquidity of the housing market from a macro perspective, and analyze the financial changes of household owners caused by housing policy from a micro perspective.

In an embodiment of the present disclosure, a Korean housing simulation model may be used. For example, as environmental and policy variables used by the model, the Korean housing market index and household financial welfare data of the National Statistical Office can be used as micro-family data of the Korean government's economic policy.

In one embodiment of the present disclosure, the structure of the housing market model may include agents, interactions, environments, and houses. In one embodiment, there are three types of model agents, which may be furniture, external housing providers, and real estate agents. Here, the household decides to purchase, sell, and lease a home as needed using savings and loan services. External housing providers are not households, but are actors in the concept encompassing participants such as governments, companies, and architects who are directly participating in the housing market, and are in charge of supplying new housing. A real estate broker receives information on a house for sale or lease from households, registers it in the market, and provides market information when requested by households.

In one embodiment, the interaction between model agents may include a house sale transaction or a lease agreement between households or with an external supplier.

In one embodiment, the model environment may include the government's housing policy, macroeconomic environment, population availability*?* access, new housing supply, and the like.

In one embodiment, the model house has characteristics such as market value and type of house, and characteristics of owners, residents, etc. may change according to a decision of the household.

In one embodiment, the housing market model may be performed by repeating the update process, the list process, and the buy process sequentially. In the update process, the characteristics of furniture and houses are updated, and new furniture is added and the housing supply process can be performed. The listing process is the process by which households decide whether to list their homes on the market. When listing, you determine the transaction type and transaction price. The listing can be made through an intermediary. In the buy-in process, households may move in through a home purchase or lease contract or purchase additional homes. In this process, households can decide on market participation, where they live, and choose their preferred housing type. Families can receive housing information that meets the conditions from a real estate agent, and can select and contract a house in consideration of the budget.

Experiment result

10,000 agents were applied to the simulation model for experimentation.

As for the characteristics of each agent, which is the microscopic result of the simulation, there are a total of 7 residential areas, savings, income, loans, residential housing type, residential housing occupancy type, and the number of houses held. Table 1 is a table showing the characteristics of a total of 10 micro households for each agent by pre-processing the occupancy type of a residence by one-hot encoding.

[Table 1] Characteristics of micro furniture generated by agent

Here, for the calculation of the matching degree of the simulation, the macroscopic results of the apartment sales in the metropolitan area, the price index of cheonsei and the transaction volume are used. When the cluster analysis handled based on the simulation microscopic results that exist for each agent for as long as the simulation time is performed, a result as shown in FIG. 3 is obtained.

FIG. 4 is a graph showing results of simulation matching using a method of calibrating micro-simulation parameters using machine learning according to an embodiment of the present disclosure.

The simulation match degree results are shown in FIG. 4. The graph 410 of FIG. 4 is a matching degree derived when experimenting with a heuristic technique, that is, an ad-hoc parameter set found by a human, and the graph 420 is derived as a result of micro parameter calibration. It is the degree of consistency. The result of FIG. 4 is a result of dividing into four clusters and performing micro parameter calibration. The corrected micro parameter is willingness to pay, which represents the maximum percentage of the current household's available budget to pay to purchase a home. Calibration can be performed on a plurality of micro parameters. As shown in FIG. 4, the matching degree derived from the micro parameter calibration result can be improved by 4.2% by fast calibration (around 10 iterations) about 10 times compared to the matching degree using the heuristic technique.

FIG. 5 is a graph of a metropolitan area apartment trading volume obtained by using a calibrated parameter based on a method for calibrating a micro-simulation parameter using machine learning according to an embodiment of the present disclosure. A graph 510 shown in FIG. 5 is a simulation result value, and a graph 520 is a result value of validation data.

6 is a graph of a metropolitan area apartment trading volume calculated using an existing ad-hoc parameter. The graph 610 of FIG. 6 is a simulation result value, and the graph 620 is a validation data result value. Comparing FIG. 5 and FIG. 6, when experimenting with the existing parameters (ad-hoc, human-adjusted parameters) (FIG. 6), the simulation result as a whole was down to the validation data, whereas the corrected parameters through machine learning. It can be seen that the simulation results obtained by substituting for the simulation input parameters better simulate the validation data.

In this disclosure, while presenting a methodology for calibration of micro-simulation parameters, a plurality of statistical models are used to obtain the reliability of the simulation model. For example, for cluster analysis, the most widely used probabilistic self-coding model among the deep generative models and the Dirichlet process mixture model, which is a nonparametric clustering model, were used, and micro parameters that derive the maximum degree of matching were used. Bayesian optimization was used to obtain, but the present disclosure is not limited thereto, and various modifications may be used by those skilled in the art.

Although the method has been described through specific embodiments, it is also possible to implement the method as computer-readable code on a computer-readable recording medium. The computer-readable recording medium includes all types of recording devices that store data that can be read by a computer system. Examples of computer-readable recording media include ROM, RAM, CD-ROM, magnetic tapes, floppy disks, optical data storage devices, and the like, and also include those implemented in the form of carrier waves (for example, transmission through the Internet). . Further, the computer-readable recording medium is distributed over a computer system connected by a network, so that codes that can be read by the computer in a distributed manner can be stored and executed. Further, functional programs, codes, and code segments for implementing the above embodiments can be easily inferred by programmers in the technical field to which the present disclosure belongs.

In the embodiments disclosed herein, the arrangement of the illustrated components may vary depending on the environment or requirements in which the invention is implemented. For example, some components may be omitted or some components may be integrated and implemented as one. Also, the arrangement order and connection of some components may be changed.

In the above, various embodiments of the present disclosure have been illustrated and described, but the present disclosure is not limited to the specific embodiments described above, and the foregoing embodiments deviate from the gist of the present disclosure as claimed in the appended claims. Without this, various modifications may be made by those skilled in the art to which the present invention pertains, and these modified embodiments should not be understood separately from the technical spirit or scope of the present disclosure. Therefore, the technical scope of the present disclosure should be defined only by the appended claims.

110: Pre-processing target data
120: graph of real data
130: ABM parameters and structure
140: two-dimensional drawing
150: Simulation result graph
160: simulation error analysis
170: micro cluster analysis
180: micro parameter estimation

Claims (5)

As a method of calibrating micro-simulation parameters using machine learning in an agent-based simulation performed by a computer,
(a) generating validation data by preprocessing moving object data, which is target data to be estimated through simulation, including characteristic data of a plurality of agents, and converting it into a predetermined data format;
(b) Agent-Based Modeling and Simulation (ABMS) is performed using scenario data for reflecting the initial parameter set and external variables for the characteristics of the plurality of agents into the simulation as ABM input data. Obtaining agent micro data of each of the plurality of agents by performing, and obtaining agent micro data by aggregating the agent micro data;
(c) performing micro-clustering analysis on micro-data of each of the plurality of agents to generate at least one clustering and performing cluster analysis; (d) analyzing an error by comparing the validation data and the agent macroscopic data;
(e) If the error exceeds a predetermined reference value,
(e-1) calibrating the parameters for each of the at least one cluster to reduce the error, and setting a correction parameter set for each of the at least one cluster;
(e-2) Each of the plurality of agents by setting the calibration initial parameter set and the scenario data for the characteristics of the plurality of agents as ABM input data, and performing agent-based modeling and simulation on the ABM input data. Obtaining agent micro data of the agent and collecting the agent micro data to obtain agent macro data;
(e-3) performing steps (d) and (e);
(f) when the error is less than or equal to a predetermined reference value, determining the corrected parameter as a final parameter
Micro-simulation parameter calibration method comprising a. As a method of calibrating micro-simulation parameters using machine learning in an agent-based simulation performed by a computer,
(a) generating validation data by preprocessing moving object data, which is target data to be estimated through simulation, including characteristic data of a plurality of agents, and converting it into a predetermined data format;
(b) Agent-Based Modeling and Simulation (ABMS) is performed using scenario data for reflecting the initial parameter set and external variables for the characteristics of the plurality of agents into the simulation as ABM input data. Obtaining agent micro data of each of the plurality of agents by performing, and obtaining agent micro data by aggregating the agent micro data;
(c) performing micro-clustering analysis on micro-data of each of the plurality of agents to generate at least one clustering and performing cluster analysis;
(d) analyzing an error by comparing the validation data and the agent macroscopic data; And
(e) if the error exceeds a predetermined reference value, estimating micro parameters
Micro-simulation parameter calibration method comprising a. The method of claim 1,
The step of performing the cluster analysis is a micro-simulation parameter, which is a step of performing cluster analysis using the latent expression as a new input value after learning the hidden representation of the agent through a variant autoencoder (VAE). Calibration method. A computer-readable recording medium containing one or more instructions,
The one or more instructions, when executed for a computer, cause the computer to perform the method of any one of claims 1 to 3. As a micro-simulation parameter calibration device using machine learning in agent-based simulation,
A validation data generation module configured to generate validation data by preprocessing moving object data, which is target data to be estimated through simulation, including characteristic data of a plurality of agents, and converting it into a predetermined data format;
Using the scenario data for reflecting the initial parameter set and external variables for the characteristics of the plurality of agents into the simulation as ABM input data, agent-based modeling and simulation (ABMS) is performed, and the A modeling module configured to obtain agent micro data of each of a plurality of agents, and to obtain agent macro data by aggregating the agent micro data;
A cluster analysis module configured to generate at least one cluster by performing micro cluster analysis on the micro data of each of the plurality of agents and perform cluster analysis;
An error analysis module configured to compare the validation data with the agent macroscopic data to analyze an error; And
A micro parameter estimation module, configured to calibrate the parameter for each of the at least one cluster to reduce the error, and set a calibration parameter set for each of the at least one cluster
Micro-simulation parameter calibration device comprising a.

Images