JP6853955B2 - People flow pattern estimation system, people flow pattern estimation method and people flow pattern estimation program - Google Patents

Description

The present invention relates to a human flow pattern estimation system for estimating a human flow pattern in an arbitrary area, a human flow pattern estimation method, and a human flow pattern estimation program.

There is a desire to find parameters for a people flow simulator that will produce data that is closest to the population distribution (people flow pattern) in a given area. Here, the "parameter" is the "number of people in the group, where they came from (departure point), what time they came from (departure time), where they are returning from (final arrival point), and where they go through. It consists of "no (waypoint), staying time of the way value". The starting point and the final reaching point may be the same.

Various methods are known as methods for obtaining such parameters.

For example, an approximate Bayesian Computation (ABC) is known. ABC is a type of Monte Carlo method, and is a method of sequentially obtaining parameters with high performance from the values of randomly selected parameters. For example, in the simplest type of ABC, pseudo data is generated from the simulator, and if the pseudo data is close to the actual data, the parameters used in the simulator are recognized as correct.

At ABC, sampling is determined based on prior knowledge and simulation is performed. Then, in ABC, only those whose results (pseudo data) fit well with the actual data are filtered.

However, in ABC, there is a problem that it is difficult to match when the parameters are high-dimensional. Therefore, ABC requires a large number of samples.

Kernel ABC utilizes all samples and expresses the posterior distribution as a kernel mean weighted by the goodness of fit (see Non-Patent Document 1). In many cases, there is an incentive to sample from the posterior distribution. However, sampling from the kernel mean of the posterior distribution is not easy.

Kernel Herding allows sampling from the kernel mean of the posterior distribution. In particular, sampling by Kernel Herding can cover the characteristics of the distribution well with a small number, and is called a super sample.

Automatic Relevance Determination (ARD) is an effective tool for removing redundancy by leaving important things from a given set of parameters.

On the other hand, various methods for estimating the flow pattern have been proposed.

For example, Patent Document 1 discloses a human flow analysis system that supports efficient operation of a city. The human flow analysis system disclosed in Patent Document 1 includes a region designation unit, a human flow calculation unit, and a display control unit. The area designation section designates the area on the map. The people flow calculation unit calculates the people flow related to the designated area of the area designation unit based on a plurality of people flow information including time and position information. The display control unit displays the flow line of the main flow of people calculated by the people flow calculation unit on the map displayed on the display unit. The human flow analysis system includes a simulation unit. The simulation unit calculates the human flow by simulation. For example, the simulation unit calculates the human flow when a new agent is generated based on the human flow database. Here, an agent is a virtual entity that is set in place of the actual environment in order to generate changes in the flow of people and traffic due to changes in the environment, such as changes in residents and the opening and closing of facilities. To say.

Further, Patent Document 2 discloses a spatiotemporal variable prediction device capable of accurately predicting the value of a spatiotemporal variable having a temporal and spatial correlation. In Patent Document 2, as time-series data of spatiotemporal variables, population distribution, speed / direction of human flow / traffic flow, reserves of mineral resources such as gold and diamond, meteorological data such as precipitation, land price, etc. are assumed. There is. In Patent Document 2, as an embodiment, the distribution of unobserved points or future spatiotemporal variables is estimated and predicted under the condition that the time series data of the population density distribution is given as the observed data.

The spatiotemporal variable prediction device disclosed in Patent Document 2 includes a learning unit. Based on the set of observation data stored in the population density information storage unit, the learning unit has multiple hyperparameters of the kernel function, which is a function that defines the similarity between observation data, and a plurality of observation data for each of the observation data. Learn the burden rate, which is a parameter representing each contribution of the Gaussian process. The learning unit includes a burden rate estimation unit, a Gaussian process parameter estimation unit, and an iterative determination unit. The burden rate estimation unit is a unit load which is a parameter representing the contribution of each of a plurality of units composed of a plurality of Gaussian processes based on a set of observation data and each hyperparameter of a kernel function of a plurality of Gaussian processes. Estimate the rate and the burden rate, which is a parameter representing the contribution of each of the multiple Gaussian processes. The Gaussian parameter estimator is based on the set of observation data, the unit burden rate of each of the multiple units, and the burden rate of each of the multiple Gaussian processes, and for each of the multiple Gaussian processes, the kernel of the Gaussian process. Estimate each hyperparameter of the function. The repetition determination unit repeats the estimation by the burden rate estimation unit and the estimation by the Gaussian process parameter estimation unit until a predetermined repetition end condition is satisfied.

Further, Patent Document 3 discloses a human flow survey support system that more efficiently executes a large-scale human flow distribution estimation process. The human flow survey support system disclosed in Patent Document 3 uses a distributed processing platform to generate a large number of hypotheses of human flow distribution based on a human flow model that expresses changes in human flow. Then, by evaluating the compatibility between each of these hypotheses and the measurement data, the human flow distribution that does not contradict the measurement data is calculated. At this time, the time evolution process of the human flow distribution based on the human flow model is calculated for each similar human flow model, so that the calculation is parallelized and duplicate calculation is eliminated. By parallelizing the processing for each human flow model, the evaluation with the measured value is speeded up by the spatiotemporal index with the measurement data.

The human flow survey support system disclosed in Patent Document 3 is a human flow distribution distribution system having a distributed processing platform that distributes calculations to a plurality of computers and a function of distributing human flow distribution data that estimates the human flow distribution from hypothetical data of the human flow distribution. And a human flow measuring instrument that measures real-time measurement data of human flow. In Patent Document 3, the distributed processing platform duplicates the entire human flow 100% survey data that matches the current time by a predetermined number of minutes, and randomly shifts the position of each person in each hypothesis to obtain human flow hypothesis data.

Further, Patent Document 4 discloses a product information providing system that supports the purchase of products in a physical store. In Patent Document 4, the flow line is analyzed and estimated from the camera or GPS (Global Positioning System) information to grasp the interest of the visitor and determine the recommended product.

Non-Patent Document 2 discloses a technical idea of analyzing migratory behavior between areas in units of 250 m mesh by utilizing mobile spatial statistics.

Non-Patent Document 3 discloses a technical idea of probabilistically estimating a stop-by place by a hierarchical migration model based on questionnaire information.

International Publication No. 2016/067369 JP-A-2017-0331198 Japanese Unexamined Patent Publication No. 2014-11292 Japanese Unexamined Patent Publication No. 2015-022354

Shigeki Nakagome, Shuhei Mano and Kenji Fukumizu (2013) Kernel Approximate Bayesian Computation in Population Genetic Inferences, Statistical Applications in Genetics and Molecular Biology, Volume 12, Issue 6, Pages 667-678 Analysis of pedestrian migration in the city center (Part 2) -Attributes of residents in the city center-, Nagoya City Center Research Report, No. 117 2015. Predictive model for detailed analysis and policy simulation of migratory behavior in Machinaka, JSCE Proceedings, Vol. 71, No.5 323-335, 2015

Prior art (ABC) has the following problems.

The first problem is that if the number of simulator parameters to be estimated is large, the accuracy of the parameters will decrease.

In addition, ABC is inefficiently searching for parameter candidates. Therefore, the second problem is that it takes a wasteful calculation time because the simulation is performed even for the parameter candidates that are not so meaningful.

Furthermore, the third problem is that when a parameter that is not actually needed is searched for with wrong knowledge, it is unconditionally accepted.

Each of the above-mentioned prior arts (Patent Documents 1 to 4 and Non-Patent Documents 1 to 3) has the following problems.

Patent Document 1 merely discloses the technical idea of calculating the human flow when a new agent is generated based on the human flow database. Therefore, in Patent Document 1, it is necessary to store each individual's outing behavior in a human flow database in advance. In other words, Patent Document 1 is a technical idea that does not hold when each individual's outing behavior is unknown in advance.

Patent Document 2 simply discloses the technical idea of estimating and predicting the unobserved point or the future spatiotemporal variable distribution under the condition that the time series data of the population density distribution is given as the observed data. Not too much. Further, Patent Document 2 describes that it can be applied as time-series data targeting the speed and direction of human flow, but specifically how to apply it to the speed and direction of human flow. Does not disclose or suggest anything.

In Patent Document 3, the distributed processing platform duplicates the entire human flow 100% survey data that matches the current time by a predetermined number of minutes, and randomly shifts the position of each person in each hypothesis to obtain human flow hypothesis data. Therefore, a huge amount of data is required for each person.

In Patent Document 4, since the information of the camera and GPS is utilized, it is possible to acquire the action trajectory of an individual. However, the technique of Patent Document 4 leads to personal identification, and therefore has a high possibility of infringing privacy.

As described above, Kernel ABC disclosed in Non-Patent Document 1 has a problem that it cannot be sampled again using kernel mean. That is, the asymptotic method cannot be completed.

Since Non-Patent Document 2 uses population distribution information such as mobile spatial statistics, it is possible to protect privacy. However, in the method disclosed in Non-Patent Document 2, it is difficult to estimate fine migratory behavior because the analysis particle size is rough.

In the method disclosed in Non-Patent Document 3, migratory behavior is analyzed based on population distribution and questionnaire information based on a probabilistic behavior model. However, such an analysis method cannot grasp the detailed relationship between facilities, and therefore the information is insufficient to utilize the analysis result for town planning.

An object of the present invention is to provide a human flow pattern estimation system and a human flow pattern estimation method that solve the above-mentioned problems.

The person flow pattern estimation system according to the present invention is a person flow pattern estimation system that estimates a person flow pattern that is a movement in a person flow group unit in the city, and defines the person flow pattern by incorporating the map information and facility information in the city. The machine learning is provided with a human flow simulator that outputs population distribution data y at each point and time from a sample of parameter θ, and a machine learning device that determines the parameter θ set in the human flow simulator by iterative learning. The apparatus includes a first sample acquisition unit that acquires N (N is an integer of 2 or more) samples θ _i (1 ≦ i ≦ N) of the _{parameter θ from the prior distribution P; and the N samples θ i. A simulation execution unit that acquires N population distribution data y i} at each point and time from the human flow simulator, respectively, by sequentially supplying the N human flow simulators; the N pieces using the kernel function k. population data y _i and each point of the similarity calculation unit for calculating a N number of similarities w _i between the correct answer data y ^* representing the true population distribution at each time; the N number of similarities A kernel average construction unit that constructs a kernel average μ of the posterior distribution from w _i , the N samples θ _i, and the kernel function k; and a parameter θ from the kernel average μ and the kernel function k. A second sample acquisition unit that acquires T (T is an integer of 2 or more) new samples θ _t (1 ≦ t ≦ T); and one sample θ selected from the _{T new samples θ t.} difference || y k _-y * ^|| between the new population distribution data y _k obtained the correct answer data y ^* from the human flow simulator by supplying _k to the pedestrian flow simulator, the predetermined threshold ε a sample evaluation unit to determine a smaller or not; wherein the sample evaluation unit, if the difference ^|| y k _-y * || determines that the smaller the predetermined threshold epsilon, said selected The sample θ _k is estimated and output as a parameter θ that defines the human flow pattern.

The people flow pattern estimation method of the present invention incorporates the map information and facility information in the city in the people flow pattern estimation system that estimates the people flow pattern, which is the movement of the people flow group in the city, and the parameter θ that defines the people flow pattern. This is a first method of estimating a human flow pattern in which the parameter θ set in the human flow simulator is determined by iterative learning using a human flow simulator that outputs population distribution data y at each point and at each time from the sample. The sample acquisition unit acquires N (N is an integer of 2 or more) samples θ _i (1 ≦ i ≦ N) of the parameter θ from the prior distribution P; the simulation execution unit obtains the N samples θ _{i. Are sequentially supplied to the human flow simulator to acquire N population distribution data y i} at each point and at each time from the human flow simulator, respectively; the similarity calculation unit uses the kernel function k to obtain the above-mentioned N number of population distribution data y _i and each point, the N to calculate the similarity w _i between the correct answer data y ^* representing the true population distribution at each time; the kernel average construction unit, the N a similarity w _i of the N number of samples theta _i, from said kernel function k, to build a kernel mean μ of the posterior distribution; second sample acquisition unit, the kernel average μ and the kernel function k From the above, T (T is an integer of 2 or more) new samples θ _t (1 ≦ t ≦ T) of the parameter θ are obtained; the sample evaluation unit selects from the _{T new samples θ t.} The difference between the _{new population distribution data y k} obtained from the human flow simulator and the correct answer data y ^* by supplying one sample θ _{k to the human flow simulator || y k −} y ^* || determining whether a predetermined smaller than the threshold value epsilon; the sample evaluation unit, the difference || y k _-y * ^{if ||} determines that the smaller the predetermined threshold epsilon, the sample theta _k of said selected Is estimated as a parameter θ that defines the human flow pattern, and is output.

According to the present invention, it is possible to estimate high-dimensional parameters of the human flow pattern, and it is possible to reduce the calculation time.

It is a block diagram which shows the structure of the person flow pattern estimation system which concerns on one Embodiment of this invention. It is a figure which shows an example of the population time distribution data which shows the correct answer data. It is a figure which shows an example of the map information and facility information (map / building data) incorporated (set) in the person flow simulator used in the person flow pattern estimation system shown in FIG. 1. It is a figure which shows an example of a person flow pattern. It is a block diagram which shows the main processing part realized by the machine learning apparatus used in the human flow pattern estimation system shown in FIG. It is a figure for demonstrating the operation of the data input part used in the machine learning apparatus shown in FIG. It is a figure for demonstrating the operation of the 1st sample acquisition part used in the machine learning apparatus shown in FIG. It is a figure for demonstrating the operation of the simulation execution part used in the machine learning apparatus shown in FIG. It is a figure for demonstrating the operation of the Kernel ABC execution part (similarity calculation part) used in the machine learning apparatus shown in FIG. It is a figure for demonstrating the operation of the kernel average construction part used in the machine learning apparatus shown in FIG. It is a figure for demonstrating the operation of the 2nd sample acquisition part used in the machine learning apparatus shown in FIG. It is a figure for demonstrating the operation of the sample evaluation part used in the machine learning apparatus shown in FIG. It is a flowchart for demonstrating one Example of the operation of the person flow pattern estimation system shown in FIG. It is a figure which shows an example of the 4 mesh model used for the experiment when the human flow pattern estimation system shown in FIG. 1 was executed. It is a figure which shows an example of 5 rules set in the experiment when the person flow pattern estimation system shown in FIG. 1 is executed. It is a figure which shows the experimental result of the 4 mesh model shown in FIG.

First, in order to facilitate the understanding of the present invention, the meanings of the terms used in the present specification will be briefly described.

The Approximate Bayesian Computation (ABC) is a type of Monte Carlo method, in which parameters with high performance are sequentially obtained from the values of randomly selected parameters. For example, ABC generates pseudo data from a simulator, and if the pseudo data is close to the actual data, the parameters used in the simulator are recognized as correct.

In other words, ABC is a non-likelihood method for Bayesian estimation that applies a rejection algorithm method to the difference between observed and simulated data.

Kernel ABC utilizes all samples and expresses the distribution as a kernel mean weighted by goodness of fit.

The kernel method is a methodology for constructing maps to higher dimensional space. The kernel method maps data into a high-dimensional reproducing kernel Hilbert space (generally infinite dimension) and transforms it into data that is easy to analyze. A Hilbert space is a vector space, a set that has an inner product and is complete.

A Hilbert space H having a function on the set Ω as an element is a reproducing kernel Hilbert space (RKHR), which means that there is φ _X ∈ H for any x ∈ Ω and any f ∈. Against H
<F, φ _X > = f (x)
Is true. φ _X is called a regenerated kernel.

Kernel Herding makes good use of the characteristics of the kernel mean distribution and selects a representative point with less fog as a super sample. In other words, Kernel Herding expresses the distribution of appropriate settings with a small number of sample points.

In ARD, by adjusting the hyperparameters, most of the values of the parameters take 0 in the sample from the prior distribution, and the extra ones are deleted. This is usually achieved by introducing a Dirichlet or beta distribution as a prior distribution. In the case of the Dirichlet distribution, if the value of the hyperparameter associated with it is made small, the sample from the distribution takes a value close to 0, and the force to explain the phenomenon with a small number of parameters works.

In other words, if you want to control the redundancy of the parameters, get a sample of the parameters from the Dirichlet distribution or the beta distribution. There are hyperparameters that control parameter redundancy in the Dirichlet distribution and beta distribution. By changing the value of the hyperparameter, the redundancy of the parameter sample is adjusted. For example, if you set hyperparameters to eliminate redundancy, only a few of the sample elements will have non-zero values and many will have zero values. As a result, the element that takes a value of 0 has no substantial effect. Thus, the resulting posterior distribution kernel mean is very redundant.

[Related technology]
In order to facilitate the understanding of the present invention, related techniques and their problems will be described.

Given the fact (truth) of the problem, we want to know a set of simulator parameters that generate data that is close enough to that fact (truth). I want to apply Bayesian rules, but the likelihood function cannot be evaluated. However, sampling is possible by the following equation based on the prior probability π (θ) from the simulator likelihood p (y | θ).

Therefore, it is conceivable to use the ABC framework. Table 1 shows the ABC algorithm.

However, ABC has the following problems.

The first problem is that when the dimension of the data is high, the parameters do not easily meet the tolerance level and the algorithm does not finish in a realistic time.

The second problem is that the higher the dimension of the parameter, the sooner the performance of ABC deteriorates.

The third problem is that it takes time to calculate the simulation, so the number of executions cannot be omitted.

[Explanation of Outline of Algorithm According to the Present Invention]
In the present invention, Kernel ABC is used to solve the first and second problems, and Kernel Herding is used to solve the third problem.

The outline of the algorithm of the present invention is as follows.
A1. Extract the set of parameters N times from the prior distribution.
A2. Data is generated N times from the simulator using a set of N parameters.
A3. Kernel ABC
-By calculating the kernel mean of the posterior distribution, the parameters of the simulator that generate data close to the observed data are inferred.
A4. Kernel Herding
• Take a useful sample T times from the kernel mean of the posterior distribution for the next iteration.
A5. To converge A2. Repeat from.

Next, Kernel ABC will be described.
Kernel ABC is based on Kernel Bayesian rules and is a nonparametric implementation of Bayesian rules.
B1. Extract the set of parameters N times from the prior distribution.
B2. By inputting a set of parameters into the simulator N times, N simulated data are acquired.
B3. Calculate the similarity (weight) between the true data and each simulated data.
B4. B1. By averaging the set of parameters in the above by their similarity (weight), the kernel average of the parameter posterior distribution is calculated according to the following formula.

Next, the kernel average will be described.
X is a random variable that takes a value on the measuring space Ω, k is a definite kernel on the measuring space Ω, and H is a reproducing kernel Hilbert space (RKHS) defined by the definite kernel k. Suppose there is. At this time, the kernel average is calculated by the following equation.

The characteristic kernel allows the probability distribution to be uniquely embedded in the reproducing kernel Hilbert space. The characteristic kernel is, for example, a Gaussian kernel or a Laplace kernel.

Next, the advantages of using kernel averaging and kernel Bayesian rules will be discussed. Bayesian rules can be applied nonparametrically without assuming a parametric model between variables or a density function. Therefore, the kernel mean representing the probability distribution is estimated nonparametrically from the data.

Next, Kernel Herding will be described.
The output of Kernel ABC is the kernel mean of the posterior distribution in the reproducing kernel Hilbert space, but it is not itself an posterior distribution. Therefore, the sampling method from the posterior distribution represented by the Markov chain Monte Carlo method cannot be used. However, Kernel Herding allows samples from posterior distributions.

Furthermore, the sample from Kernel Herding is called a super sample that well reflects the characteristics of the distribution.

Kernel Herding can be interpreted as sampling by greedily minimizing the following error.

Kernel Herding takes samples from areas of high density in the distribution. That is, it is considered to search for the mode of distribution. Also, away from the area where the sample was taken once, the next sample is taken.
The above can be summarized as follows.

Kernel ABC can handle relatively high dimensional parameters, can handle relatively high dimensional data (simulated data / true problem data), and has manageable computational complexity.

Kernel Herding can reduce the number of simulations required.

[Explanation of Outline of Invention]
Application to human flow simulator In the first place, data assimilation of human flow simulator regarding time series of population distribution such as mobile spatial statistics has not been attempted so far. At that time, there are the following problems, and the proposed method according to the present invention addresses them.

1) The human flow simulator is a multi-agent simulator, and in many cases, the simulation takes a very long time to calculate.

On the other hand, in the proposed method, data assimilation can be achieved with a smaller number of simulations by using Kernel herding. This is because kernel herding efficiently obtains samples of important parameters with more information.

In addition, the error converges with O (1 / T). On the other hand, in the parameter sample acquisition by the Monte Carlo method or the like, the slower O (1 / √T) is obtained.

2) Although it is not a human-style simulator, regarding other data assimilation, it is known that Bayesian estimation using ABC etc. does not work well when the parameter dimension is high.

On the other hand, in the proposed method, Kernel ABC supports high-dimensional parameters by performing Bayesian inference in the reproducing kernel Hilbert space.

3) If the prior knowledge is wrong, the prior distribution may be constructed by assuming a flow pattern that does not exist in a certain area. For example, in Ginza, there is a popular pattern such as having lunch at the Yoshinoya, buying a watch at Rolex, buying a car at Ferrari, and eating dinner at Matsuya. In the related technology, there was no mechanism to automatically delete such a flow pattern that is unlikely to exist by an algorithm.

On the other hand, in the proposed method, ARD introduces sparsity to delete such patterns.

Regarding the algorithm, the proposed method makes it possible to perform point estimation when the likelihood function cannot be evaluated.

In related techniques, Bayesian inference could be performed when the likelihood function could not be evaluated (ABC, etc.). In other words, it was possible to obtain the correct posterior distribution. However, in ABC, point estimation in a situation where there is one true value cannot be performed when the likelihood function cannot be evaluated.

On the other hand, in the proposed method, the sample from the posterior distribution acquired by Kernel ABC + Kernel Herding is regarded as the sample from the prior distribution, and point estimation is achieved by repeatedly applying Kernel ABC + Kernel herding. ing.

In the first application, the posterior distribution expressed by the following formula can be obtained.

これは、

ということであり、Ｎ回目の適用では、

となる。つまり、尤度関数ｌ(ｙ｜θ) がＮ乗され、影響が強くなり、一方で事前分布π(θ) の影響が少なくなっている。これを無限回繰り返したあと、Ｐ(θ｜ｙ)を最大化するθを取得することは、π(θ)の影響が消えているため、ｌ(ｙ｜θ)を最大化するθを取得することと一致すると考えられる。つまり、無限回のパラメータに関するベイズ推定を同一のデータに対して繰り返し、Ｐ(θ｜ｙ)を最大化するθを取得することは点推定の最尤法と一致すると考えることができる。これを踏まえ、尤度関数が評価できないときにベイズ推定を行う手法であるABCを使うことで、尤度関数なしの最尤法を実行できる。これをRecursive ABC-MLEと名づけた。このアルゴリズムの内部でkernel ABCを本明細書では用いているが、尤度関数無しに事後分布を求めることができる手法であればどれでも構わない。また、このアルゴリズム内部でＰ(θ｜ｙ)を最大化するθを取得することは、Kernel Herdingのmode-seeking（最頻値追跡）という特性により可能となっている。しかしながら、取得する手法は、必ずしもKernel Herdingである必要はなく、最頻値をサンプリングする特性のある手法であればよい。 When this P (θ | y) is used as a prior distribution and applied for the second time, the following formula is obtained.

this is,

So, in the Nth application,

Will be. That is, the likelihood function l (y | θ) is raised to the Nth power, and the influence becomes stronger, while the influence of the prior distribution π (θ) becomes smaller. After repeating this infinitely, acquiring θ that maximizes P (θ | y) eliminates the influence of π (θ), so acquiring θ that maximizes l (y | θ). It is considered to be consistent with what you do. That is, it can be considered that the Bayesian estimation for the infinite number of parameters is repeated for the same data to obtain θ that maximizes P (θ | y), which is consistent with the maximum likelihood method of point estimation. Based on this, the maximum likelihood method without the likelihood function can be executed by using ABC, which is a method for performing Bayesian estimation when the likelihood function cannot be evaluated. This was named Recursive ABC-MLE. Although kernel ABC is used in this algorithm inside this algorithm, any method that can obtain the posterior distribution without a likelihood function may be used. Further, it is possible to acquire θ that maximizes P (θ | y) inside this algorithm due to the mode-seeking (mode-seeking) characteristic of Kernel Herding. However, the method to be acquired does not necessarily have to be Kernel Herding, and any method having a characteristic of sampling the mode may be used.

Kernel Herding efficiently obtains a sample of important parameters with more information from the posterior distribution. Regarding this, the error converges with O (1 / T).

On the other hand, when the above Bayesian estimation is repeated by acquiring a parameter sample by a Markov chain Monte Carlo method or the like, which is a method of a related technique, the result is slower O (1 / √T).

The essence of this algorithm is to enable maximum likelihood when the likelihood function cannot be evaluated. As the application destination of this algorithm, not only the estimation of the human flow pattern described in detail later but also the following can be considered.

1) Estimate the parameters of the alpha stationary distribution, which is often used to properly model the variance (volatility) of price fluctuations in financial time series such as stock prices, according to the data.

2) Estimate the parameters of the model that describes the time evolution of the population of organisms such as flies according to the data.

3) In population genetics, estimate model parameters (mutation rate, etc.) that realize the genetic diversity of given data.

Even if this algorithm is a software-related invention, it has the effect of reducing the calculation time, so that it improves (improves) computer-related technology and does not fall under the abstract concept.

[Embodiment]
FIG. 1 is a block diagram showing a configuration of a human flow pattern estimation system 100 according to an embodiment of the present invention.

The illustrated human flow pattern estimation system 100 includes a data processing device 200 for processing data, a storage device 300 for storing programs and data described later, an input device 400 for inputting data, and an output device 500 for outputting data. It has.

The output device 500 includes a display device such as an LCD (Liquid Crystal Display) or a PDP (Plasma Display Panel) or a printer. The output device 500 has a function of displaying various information such as an operation menu and printing out the final result in response to an instruction from the data processing device 200.

The storage device 300 includes a memory such as a hard disk, a read-only memory (ROM), and a random access memory (RAM). The storage device 300 has a function of storing processing information 310 and a program 320 required for various processes in the data processing device 200.

The data processing device 200 includes a microprocessor such as an MPU (micro processing unit) and a central processing unit (CPU). The data processing device 200 has a function of reading the program 320 from the storage device 300 and realizing various processing units that process data according to the program 320.

The data processing device 200 includes a human flow simulator 210 and a machine learning device 220.

In the illustrated human flow pattern estimation system 100, it is assumed that the data required for the algorithm and the data of the population distribution in the area and time at which the human flow is to be estimated, such as the device mobile spatial statistics, are given. This is called correct answer data y ^*.

FIG. 2 is a diagram showing an example of population time distribution data showing ^{correct answer data y *.} The population time distribution data shows the population distribution for each time T = ..., t-1, t, t + 1, t + 2, .... The time T is, for example, a time every 30 minutes. Such population time distribution data can be acquired by, for example, statistical information or SNS (Social Networking Service). Here, SNS is a membership-based online service that provides various functions for maintaining and promoting social connections between people. The storage device 300 stores the correct answer data y ^* as the processing information 310.

The people flow pattern estimation system 100 is a system that estimates the people flow pattern, which is the movement of each person flow group in the city.

The human flow simulator 210 can simulate a human flow pattern in any area and time.

More specifically, the human flow simulator 210 incorporates information on population distribution at each point and time, as well as map information and facility information in the city.

FIG. 3 is a diagram showing an example of map information and facility information (map / building data) incorporated (set) in the human flow simulator 210. The storage device 300 stores map information and facility information as processing information 310. These map information and facility information are set in the human flow simulator 210 as described later.

The human flow simulator 210 outputs the population distribution data y at each point and at each time from a sample of the parameter θ that defines the human flow pattern, which is supplied from the machine learning device 220 as described later. Note that the population distribution data y obtained from the human flow simulator 210 usually does not match the ^{correct answer data y *.} This is because no one knows the flow pattern that satisfies the ^{correct answer data y *.}

FIG. 4 shows an example of a human flow pattern. The flow pattern shown in FIG. 4 shows the flow pattern after 9:00 time at points A, B, C, D, E, F, and G.

The parameter θ includes information on the starting point, departure time, waypoint, staying time, destination, and number of people of each group.

The machine learning device 220 determines the parameter θ set in the human flow simulator 210 by iterative learning as described later.

FIG. 5 is a block diagram showing a main processing unit realized by the machine learning device 200.

The machine learning device 200 includes a data input unit 221, a first sample acquisition unit 222, a simulation execution unit 223, a Kernel ABC execution unit 224, a kernel average construction unit 225, and a second sample acquisition unit 226. It consists of a sample evaluation unit 227. The Kernel ABC execution unit 224 functions as a similarity calculation unit as described later.

Next, the operation of each processing unit will be described with reference to FIGS. 6 to 12.

The operation of the data input unit 221 will be described with reference to FIG.

As shown in FIG. 6, the data input unit 221 inputs the information of the area for which the human flow pattern is to be estimated into the human flow simulator 210. More specifically, the storage device 300 stores the map information and the facility information as the area information in the processing information 310 as described above. The data input unit 221 reads map information and facility information from the storage device 300, and inputs (sets) the information to the human flow simulator 210.

In this example, the data input unit 221 reads out the map information and the facility information stored in the storage device 300 and sets them in the human flow simulator 210, but the map information and the facility supplied from the input device 400. The information may be set directly in the human flow simulator 210.

The operation of the first sample acquisition unit 222 will be described with reference to FIG. 7. The first sample acquisition unit 222 is also referred to as a "sample acquisition unit from prior distribution". ARD is involved in the first sample acquisition unit 22.

As shown in FIG. 7, the storage device 300 stores the prior distribution P, the hyperparameter η of the prior distribution P, and the number N of samples from the prior distribution P as the processing information 310. The prior distribution P is an appropriate prior distribution that can reflect the case where the sum of arbitrary parameter samples is 1 and the structure of the mixture model or the solution is the same as that of the mixture model.

By adjusting the hyperparameter η of this prior distribution P, it is determined whether or not sparsity is included in the final solution of this algorithm.

The first sample acquisition unit 222 extracts a sample from the prior distribution as shown in the following mathematical formula.

Therefore, the first sample acquisition unit 222 acquires N samples {θ _i } (i = 1 to N) of parameters to the human flow simulator 210. The N samples {θ _i } of the acquired parameters are stored in the storage device 300 as processing information 310.

The operation of the simulation execution unit 223 will be described with reference to FIG.

As shown in FIG. 8, the storage device 300 stores the above-mentioned number N of samples and N sample {θ _i } of parameters as processing information 310.

The simulation execution unit 223 _{inputs the parameter sample {θ i} } into the human flow simulator 310 N times, _{and obtains N population distribution data {y i} } (1 ≦ i ≦) from the human flow simulator 310. The obtained N population distribution data {yi} are stored in the storage device 300 as processing information 310.

The operation of the Kernel ABC execution unit (similarity calculation unit) 224 will be described with reference to FIG.

As shown in FIG. 9, the storage device 300 processes the above-mentioned N population distribution data {y _i }, the above-mentioned correct answer data y ^* , the kernel function k, and the regularization term δ. It is saved as.

ここで、Ｇはグラム行列を示し、Ｉは単位行列を示す。この得られたＮ個の類似度{ｗ_ｉ}は、処理情報３１０として記憶装置３００に保存される。 Kernel ABC execution unit 224 performs the calculation expressed by the following equation, the similarity between the solution data and the simulated population distribution data {y _i} {w _i} get (1 ≦ i ≦ N).

Here, G represents a Gram matrix and I represents an identity matrix. The resulting N number of similarities {w _i} is stored in the storage unit 300 as the processing information 310.

In this example, as the similarity calculation unit, is used a Kernel ABC execution unit 224 for calculating the degree of similarity w _i by Kernel ABC algorithm, the present invention is not limited thereto. In other words, the algorithm used by the similarity calculation unit may whatever the N similarity algorithm for outputting a {w _i} so build the kernel average of the posterior distribution, which will be described later.

The operation of the kernel average construction unit 225 will be described with reference to FIG.

As shown in FIG. 10, the storage device 300 includes a sample {theta _i} of N parameters described above, and the N similarities described above {w _i}, process information, and a kernel function k described above It is stored as 310.

この得られたカーネル平均μは、処理情報３１０として記憶装置３００に保存される。 The kernel average construction unit 225 performs the calculation represented by the following formula to obtain the kernel average μ of the posterior distribution.

The obtained kernel average μ is stored in the storage device 300 as processing information 310.

The operation of the second sample acquisition unit 226 will be described with reference to FIG. The second sample acquisition unit 226 is also referred to as a "sample acquisition unit from the posterior distribution". The second sample acquisition unit 226 executes the above-mentioned Kernel Herding.

As shown in FIG. 11, the storage device 300 stores the above-mentioned kernel average μ, the above-mentioned kernel function k, and the number of samples T to be taken out as processing information 310.

この得られた新たなパラメータのＴ個のサンプル{θ_ｔ}は、処理情報３１０として記憶装置３００に保存される。 The second sample acquisition unit 226 performs a calculation that minimizes the error appearing in the following mathematical formula _{to obtain T samples {θ t} } (1 ≦ t ≦ T) of new parameters.

The T samples {θ _t } of the obtained new parameters are stored in the storage device 300 as processing information 310.

Finally, the operation of the sample evaluation unit 227 will be described with reference to FIG.

As shown in FIG. 12, in the storage device 300, T samples {θ _t } of the above-mentioned new parameters, the above-mentioned correct answer data y ^*, and the above-mentioned threshold value ε are stored as processing information 310. doing.

すなわち、サンプル評価部２２７は、新たな人口分布データｙ_ｋと正解データｙ^＊との差分｜｜ｙ_ｋ−ｙ^＊｜｜が、閾値εより小さいか否かを判断する。上記式が満たされていれば、サンプル評価部２２７は、選択したサンプルθ_ｋを推定された人流パターンを規定する真のパラメータとして推定し出力し、処理を終了する。 First, the sample evaluation unit 227 selects an appropriate sample θ _{k from T samples {θ t} } with new parameters, and supplies the selected sample θ _k to the human flow simulator 210. As a result, the sample evaluation unit 227 acquires _{new population distribution data y k from the human flow simulator 210.} Subsequently, the sample evaluation unit 22 determines whether or not the following equation is satisfied.

That is, the sample evaluation unit 227, the difference _|| y k ^-y * || with the new population distribution data _{y k} correct answer data ^{y *} is, determines whether the difference is less than a threshold value epsilon. If the above equation is satisfied, the sample evaluation unit 227 _{estimates and outputs the selected sample θ k} as a true parameter that defines the estimated human flow pattern, and ends the process.

On the other hand, if the above equation is not satisfied, the operation returns to the operation in the simulation execution unit 223.

Alternatively, the operation returns to the operation of the first sample acquisition unit 222. At this time, if the sparsity is not sufficiently contained, the hyperparameter η which is the input of the first sample evaluation unit 222 is also changed (ARD).

Next, an embodiment of the operation of the human flow pattern estimation system 100 will be described with reference to the flowchart of FIG. 13 with reference to a specific example.

First, the data input unit 221 sets the map information and the facility information in the human flow simulator 210 (step S101).

Specifically, the data input unit 221 gives a map of Ginza or the like to the human flow simulator 210.

Next, the first sample acquisition unit 222 extracts a sample from the prior distribution P based on the hyperparameter η, _{and acquires N samples {θ i} } of the parameters to be input to the human flow simulator 210 (step S102). ..

Specifically, the first sample acquisition unit 222 fetches, for example, 1000 samples of the parameter θ that controls the population to be simulated in the Ginza map (N = 1000).

For example, one parameter θ is composed of the following elements. Considering a group of salaried workers visiting Ginza, the parameter θ is the number of people in the group, the time when they visit Ginza, the time when they leave Ginza, the location information of the store they stop at, the time spent at that store, etc. is there. The hyperparameter η can remove extra parameters by adjusting it. For example, it can be estimated that there will be no group in Ginza that only stops at convenience stores.

Subsequently, the simulation execution unit 223 _{inputs the parameter sample {θ i} } into the human flow simulator 210 N times to obtain N population distribution data {y _i } (step S103).

Specifically, the simulation execution unit 223 performs a simulation with respect to the sample of each parameter obtained in step S102. The difference is that in one sample of the parameter set, the population may be 1000, and in another sample it may be 2000. Since the set of 1000 parameters has been fetched in step S002, the simulation execution unit 223 performs the simulation 1000 times. Then, the simulated population distribution data y _i is acquired from the human flow simulator 210.

Next, Kernel ABC execution unit (similarity calculation unit) 224 performs the calculation of Equation 10 to obtain the similarity between the solution data y ^* and the simulated population distribution data {y _i} {w _i} (Step S104).

Specifically, the Kernel ABC execution unit 224 compares _{1000 differently simulated data y i} in Ginza with the correct answer data y ^{* in Ginza.} This comparison similarity w _i of the data y _i from the correct answer data y ^* and the simulation is computed. Kernel ABC execution unit 224, on the basis of this similarity w _i, to determine the plausibility of the sample θ _i of the parameters that are tied to simulated data.

Subsequently, the kernel average construction unit 225 calculates the kernel average μ of the posterior distribution according to the above equation 11 (step S105).

Specifically, the kernel average construction unit 225, based on the calculated degree of similarity w _i at step S104, with respect to samples theta _i of parameters to build the kernel average μ of the posterior distribution. This is, on the basis of the sample θ _i from the prior distribution, and re-adjust the weights use the similarity w _i, equivalent to construct a probability distribution on the reproducing kernel Hilbert space.

Next, the second sample acquisition unit 226 acquires T samples {θ _t } of new parameters by performing a calculation that minimizes the error of the above equation (12) (step S106).

Specifically, the first sample acquisition unit 226 performs sampling from the kernel average μ of the posterior distribution. At this time, a new parameter sample θ _t is sampled, but unlike _{the parameter sample θ i} acquired in step S102, a parameter sample that produces correct answer data y ^{* is required.}

Subsequently, the sample evaluation unit 227 evaluates T samples {θ _t } of the new parameter (step S107).

Specifically, the sample evaluation unit 227 _{appropriately selects one parameter θ k from the parameter set {θ t} } (1 ≦ t ≦ T) obtained in step 106, and inputs it to the human flow simulator 210. And acquire the data y _k. Then, the sample evaluation unit 227 evaluates the certainty of the parameters by comparing _{the acquired data y k} with the correct answer data y ^*. If the selected parameter θ _k is sufficiently certain, the sample evaluation unit 227 uses it as the estimated value of human flow in Ginza (step S108).

The parameter θ _k contains information on what kind of group comes from, where they stay, and where they return from Ginza. If the certainty of the parameter θ _k is not satisfactory (No in step S107), the human flow pattern estimation system 100 repeats the operation from step S103 or step S102 once again.

Next, the experimental results when the human flow pattern estimation system 100 is executed will be described.

Here, as shown in FIG. 14, an experiment of a 4 mesh model in Ginza was set. In this example, it is divided into 4 meshes (2 × 2), but in the actual example, it should be noted that it is divided into 5 × 5 = 25 meshes, for example. In the illustrated example, the four meshes consist of regions (0,0), (0,1), (1,0), and (1,1). The time was 900 to 17:00 with an interval of 30 minutes.

Five rules were set as shown in FIG. Therefore, the dimension of the estimation parameter is 40 dimensions according to the (Start, Goal, Transit × 2, Stay × 2, T, N) × 5 rule. Here, T indicates the departure time of the group, and N indicates the number of people in the group. Note that in each of these rules, N is different from the number of samples N of the parameter θ, and T is different from the number of samples T of the new parameter.

FIG. 16 shows the experimental results of the 4 mesh model of Ginza shown in FIG. In FIG. 16, the figure of Kernel ABC on the left side shows the change in the population distribution in each mesh when only the simulation execution unit 223 and the Kernel ABC execution unit 224 are executed once. The middle Kernel Herding (1 ^st ) diagram shows each mesh when the simulation execution unit 223, the Kernel ABC execution unit 224, the kernel average construction unit 225, and the second sample acquisition unit 226 are executed only once. It shows changes in population distribution. The figure of Kernel Herding (2nd) on the right shows the population distribution in each mesh when the simulation execution unit 223, Kernel ABC execution unit 224, kernel average construction unit 225, and second sample acquisition unit 226 are executed twice. Shows a change in.

From FIG. 16, it can be seen that the estimation result with the population distribution closer to the correct answer data can be obtained by (Kernel ABC + Kernel Herding) than by Kernel ABC alone. In addition, it can be seen that the estimation result having the population distribution closer to the correct answer data can be obtained by executing (Kernel ABC + Kernel Herding) twice than by executing (Kernel ABC + Kernel Herding) once.

In the above embodiment, each part of the machine learning device 220 can be realized by using a combination of hardware and software. In the form of combining hardware and software, a human flow pattern estimation program is stored in the program memory 320, and each part is realized as various means by operating the hardware of the machine learning device 220 based on the human flow pattern estimation program. To do. Further, the human flow pattern estimation program may be recorded on a recording medium and distributed. The human flow pattern estimation program recorded on the recording medium is read into the memory via wired, wireless, or the recording medium itself, and operates the processing unit or the like. Examples of recording media include optical disks, magnetic disks, semiconductor memory devices, hard disks, and the like.

To explain the above embodiment in another expression, the computer operated as the machine learning device 220 is based on the human flow pattern estimation program stored in the program memory 320, and the data input unit 221 and the first sample acquisition unit 222. This can be achieved by operating as the simulation execution unit 223, the Kernel ABC execution unit (similarity calculation unit) 224, the kernel average construction unit 225, the second sample acquisition unit 226, and the sample evaluation unit 227.

Although embodiments and examples of the present invention have been described above with reference to the drawings, those skilled in the art can use other similar embodiments and examples, and from the present invention. It should be noted that the form can be changed or added as appropriate without deviation.

The present invention is applicable to the field of estimating parameters by the maximum likelihood method when the likelihood function cannot be evaluated. For example, the present invention is also used in the field of estimating model parameters that represent the variance of financial time series, model parameters that describe the time evolution of an organism's population, and model parameters that realize genetic diversity. Applicable.

100 Human flow pattern estimation system 200 Data processing device 210 Human flow simulator 220 Machine learning device 221 Data input unit 222 First sample acquisition unit (sample acquisition unit from prior distribution)
223 Simulation execution unit 224 Kernel ABC execution unit (similarity calculation unit)
225 Kernel average construction part 226 Second sample acquisition part (sample acquisition part from posterior distribution)
227 Sample evaluation unit 300 Storage device 310 Processing information 320 Program 400 Input device 500 Output device

Claims (11)

It is a people flow pattern estimation system that estimates the people flow pattern, which is the movement of each person flow group in the city.
A people flow simulator that incorporates map information and facility information in the city and outputs population distribution data y at each point and time from a sample of the parameter θ that defines the people flow pattern.
A machine learning device that determines the parameter θ set in the human flow simulator by iterative learning, and
With
The machine learning device
A first sample acquisition unit that acquires N (N is an integer of 2 or more) samples θ _i (1 ≦ i ≦ N) of the parameter θ from the prior distribution P, and
By sequentially supplying the N samples θ _i to the human flow simulator, a simulation execution unit for acquiring _{N population distribution data y i} at each point and time from the human flow simulator, respectively.
Using a kernel function k, said N population distribution data y _i and each point, the N similarity calculating the similarity w _i between the correct answer data y ^* representing the true population distribution at each time Calculation part and
The N number of similarities w _i, wherein the N samples theta _i, from said kernel function k, and the kernel average construction unit for constructing a kernel mean μ of the posterior distribution,
A second sample acquisition unit that acquires T (T is an integer of 2 or more) new samples θ _t (1 ≦ t ≦ T) of the parameter θ from the kernel average μ and the kernel function k.
Between the new population distribution data y _k obtained from the human flow simulator and the correct answer data y ^* by supplying one sample θ _k selected from the T new samples θ _t to the human flow simulator. difference _|| y k ^-y * _|| is, a sample evaluation unit to determine a smaller or not a predetermined threshold epsilon,
With
The sample evaluation unit, if the difference ^|| y k _-y * || determines that the smaller a predetermined threshold epsilon, the sample theta _k of said selected estimates as a parameter theta defining the pedestrian flow pattern , A human flow pattern estimation system characterized by outputting. The human flow pattern estimation system according to claim 1, wherein the first sample acquisition unit acquires the N samples θ _{i while adjusting the hyperparameter η of the prior distribution P.} The similarity calculation unit consists Kernel ABC execution unit for calculating the N similarities w _i by Kernel ABC algorithm, pedestrian flow pattern estimation system according to claim 1 or 2. The Kernel ABC execution unit uses the kernel function k, the Gram matrix G, and the normalization term δ to generate the N numbers between the N _{population distribution data y i} and the correct answer data y ^*. the similarity is calculated w _i, pedestrian flow pattern estimation system according to claim 3. The kernel average building unit, based on the N samples theta _i, reproducing the N readjusting similarity w _i of the weight using a kernel function k, the kernel average μ representing a probability distribution The human flow pattern estimation system according to any one of claims 1 to 4, which is constructed on a nuclear Hilbert space. The second sample acquisition unit performs a calculation that minimizes the error obtained between the kernel function k, the new sample θ _{t, the} number of samples T to be taken out, and the kernel average μ, and performs the new calculation. The human flow pattern estimation system according to any one of claims 1 to 5, which obtains a sample θ _t. In the sample evaluation unit, wherein when the difference ^|| y k _-y * || is determined to greater than said predetermined threshold epsilon, the simulation executing unit, the similarity calculation unit, the kernel average construction unit, the The human flow pattern estimation system according to any one of claims 1 to 6, wherein the operations of the sample acquisition unit 2 and the sample evaluation unit are repeated. In the sample evaluation unit, wherein when the difference ^|| y k _-y * || is determined to greater than said predetermined threshold epsilon, the first sample acquisition unit, the simulation executing unit, the similarity calculation unit The human flow pattern estimation system according to any one of claims 1 to 6, wherein the operations of the kernel average construction unit, the second sample acquisition unit, and the sample evaluation unit are repeated. The human flow pattern estimation according to claim 8, wherein if it is desired to control the redundancy of the parameter θ before performing the repeating operation, the first sample acquisition unit adjusts the hyperparameter η of the prior distribution P. system. In the people flow pattern estimation system that estimates the people flow pattern, which is the movement of the people flow group in the city, the map information and facility information in the city are incorporated, and from the sample of the parameter θ that defines the people flow pattern, at each point and at each time. This is a method of estimating a human flow pattern in which the parameter θ set in the human flow simulator is determined by iterative learning using a human flow simulator that outputs the population distribution data y of the above.
The first sample acquisition unit acquires N (N is an integer of 2 or more) samples θ _i (1 ≦ i ≦ N) of the parameter θ from the prior distribution P.
The simulation execution unit _{sequentially supplies the N samples θ i} to the human flow simulator, thereby acquiring _{N population distribution data y i} at each point and time from the human flow simulator, respectively.
Using the kernel function k, the similarity calculation unit uses the kernel function k to provide N similarity w between the N _{population distribution data y i} ^{and the correct answer data y *} representing the true population distribution at each point and time. Calculate _{i and}
Kernel Mean construction unit is constructed the N number of similarities w _i, wherein the N samples theta _i, from said kernel function k, the kernel mean μ of the posterior distribution,
The second sample acquisition unit acquires T (T is an integer of 2 or more) of new samples θ _t (1 ≦ t ≦ T) of the parameter θ from the kernel average μ and the kernel function k.
_{The new population distribution data y k} and the correct answer data y k obtained from the human flow simulator by the sample evaluation unit supplying one sample θ _k selected from the T new samples θ _{t to the human flow simulator.} ^* difference _|| y k ^-y * || between is determined whether smaller or not a predetermined threshold epsilon,
The sample evaluation unit, if the difference ^|| y k _-y * || determines that the smaller a predetermined threshold epsilon, the sample theta _k of said selected estimates as a parameter theta defining the pedestrian flow pattern , A method for estimating a human flow pattern, which is characterized by outputting. In the people flow pattern estimation system that estimates the people flow pattern, which is the movement of the people flow group in the city, the map information and facility information in the city are incorporated, and from the sample of the parameter θ that defines the people flow pattern, at each point and at each time. This is a human flow pattern estimation program that causes a computer to determine the parameter θ set in the human flow simulator by iterative learning by using a human flow simulator that outputs the population distribution data y of the above.
The first sample acquisition procedure for acquiring N (N is an integer of 2 or more) samples θ _i (1 ≦ i ≦ N) of the parameter θ from the prior distribution P, and
A simulation execution procedure for acquiring _{N population distribution data y i} at each point and time from the human flow simulator by sequentially supplying the N samples θ _{i to the human flow simulator.}
Using a kernel function k, said N population distribution data y _i and each point, the N similarity calculating the similarity w _i between the correct answer data y ^* representing the true population distribution at each time Calculation procedure and
The N number of similarities w _i, wherein the N samples theta _i, from said kernel function k, and kernel average building procedure for building the kernel average μ of the posterior distribution,
A second sample acquisition procedure for acquiring T (T is an integer of 2 or more) new samples θ _t (1 ≦ t ≦ T) of the parameter θ from the kernel average μ and the kernel function k.
Between the new population distribution data y _k obtained from the human flow simulator and the correct answer data y ^* by supplying one sample θ _k selected from the T new samples θ _t to the human flow simulator. difference _|| y k ^-y * _|| is, a sample evaluation procedure to determine a smaller or not a predetermined threshold epsilon,
To execute,
The sample evaluation procedure, the computer, if the difference ^|| y k _-y * || determines that the smaller a predetermined threshold epsilon, the sample theta _k of said selected, parameters defining the pedestrian flow pattern A human flow pattern estimation program characterized by estimating as θ and outputting it.

Images