Prediction device, prediction method, and prediction program

Abstract

To provide a prediction device, a prediction method, and a prediction program for predicting traffic volume by a method involving a clear mechanism and low calculation cost.SOLUTION: A prediction device 10 includes a calculation unit 131 and a prediction unit 133. The calculation unit 131 calculates a probability distribution of traffic volume for each point based on a Bayesian network in which a point represented on a map is a node and a connection between points is a link. The prediction unit 133 predicts traffic volume between points based on the probability distribution.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a prediction device, a prediction method, and a prediction program. [Background technology]

[0002] In CPS (Cyber ​​Physical Systems), it is important to know how to predict what will happen in cyberspace based on information measured in the real world. For example, by measuring and predicting the flow of people and traffic using cameras and other sensors, it is expected that this will improve mobility in areas and contribute to the elimination of traffic congestion.

[0003] A common conventional method for predicting traffic volume is to use deep learning.

[0004] Also, there is known a technology for determining origin-destination traffic volume that originates from a certain point in a road network, travels a certain route, and disappears at another point, in which the traffic volume set for the departure link is distributed at each branch, and the generated traffic volume of each link is estimated using Bayesian estimation based on the through traffic volume of each link obtained by simulation (see, for example, Patent Document 1). [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Application Laid-Open No. 2002-049984 Summary of the Invention [Problem to be solved by the invention]

[0006] However, with conventional techniques, it is sometimes difficult to predict traffic volume using a method that has a clear mechanism and low computational costs.

[0007] For example, prediction methods using deep learning have the problem that the mechanism that leads to the results is unclear and that ensuring prediction accuracy requires high calculation costs. Also, the method of analyzing the principles through simulation described in Patent Document 1 uses Bayesian estimation as part of the calculation when estimating traffic volume, but there are thought to be more effective ways of using Bayesian estimation when estimating traffic volume.

[0008] The present invention has been made in consideration of the above, and aims to provide a prediction device, a prediction method, and a prediction program that can predict traffic volume using a method with a clear mechanism and low computational costs. [Means for solving the problem]

[0009] In order to solve the above-mentioned problems and achieve the object, a prediction device according to the present invention includes a calculation unit and a prediction unit. The calculation unit calculates a probability distribution of traffic volume for each point based on a Bayesian network constructed by regarding map data including nodes that are points and links connecting the nodes as a directed graph. The prediction unit predicts traffic volume between points based on the probability distribution. [Effects of the Invention]

[0010] According to the present invention, traffic volume can be predicted using a method with a clear mechanism and low calculation costs. [Brief explanation of the drawings]

[0011] [Figure 1] FIG. 1 is a diagram illustrating a model according to an embodiment. [Figure 2] FIG. 2 is a block diagram illustrating an example of the configuration of the prediction device according to the embodiment. [Figure 3] FIG. 3 is a diagram illustrating a method for synthesizing probability distributions according to the embodiment. [Figure 4] FIG. 4 is a diagram illustrating a method for updating a probability distribution according to the embodiment. [Figure 5]FIG. 5 is a flowchart showing the flow of the prediction process according to the embodiment. [Figure 6] FIG. 6 is a flowchart showing the flow of the update process according to the embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0012] Hereinafter, embodiments of a prediction device, a prediction method, and a prediction program disclosed in the present application will be described in detail with reference to the accompanying drawings. Note that the present invention is not limited to the embodiments described below.

[0013] First, a model used in the embodiment will be described with reference to Fig. 1. Fig. 1 is a diagram illustrating a model according to the embodiment. The prediction device according to the embodiment can predict traffic volume using the model described in Fig. 1.

[0014] Figure 1 shows a directed graph with nodes A, B, C, D, E, F, and G. Digital maps, including dynamic maps, are composed of data in which each branch point is represented as a node, and the connection from each point to the previous or next point is represented as a link. This representation of map data is essentially a directed graph, so the map data is considered to be a directed graph.

[0015] A directed graph can be likened to a network model of pedestrian and traffic flows. Pedestrian and traffic flows are correlated by road connections, and at each point or connection, factors such as traffic lights and stores affect branching and congestion.

[0016] Therefore, in the embodiment, a model for predicting traffic volume is created by treating the directed graph as a Bayesian network, thereby obtaining a model for predicting pedestrian and traffic flows quickly from a small amount of information.

[0017] In an embodiment, points correspond to nodes in a Bayesian network, and connections between points correspond to links in the Bayesian network.

[0018] A Bayesian network is a directed acyclic graph in which each of multiple events is a node, and the presence or absence of causal relationships between events is expressed by the links between the nodes. For example, each node in a Bayesian network is assigned a probability distribution according to the corresponding event. This makes it possible to calculate the probability (conditional probability) for each link.

[0019] In the following explanation, traffic is a concept that includes not only moving bodies such as vehicles, but also the movement of pedestrians (people flow). Furthermore, entities that generate traffic, such as moving bodies and people, are called traffic entities.

[0020] The prediction device uses the model to calculate the traffic volume at each node from the measurement results of the movement frequency of the traffic agents at each node. Furthermore, the prediction device estimates the probability that the traffic agents will move through each node from the measurement results ((1), (2), (3), and (4) in Figure 1).

[0021] Here, the prediction device calculates the probability that the traffic entity moves from node D to node F based on the dependency relationships (links) of each node ((5) in FIG. 1).

[0022] The prediction device can then predict the traffic volume entering the facility (node ​​F) from the calculation results.

[0023] It should be noted that the prediction device does not need to calculate the traffic volume and probability for all nodes and links. For example, the prediction device calculates the traffic volume for nodes corresponding to locations where the latest information can be obtained using cameras or the like, and also calculates the probability for links related to those nodes.

[0024] The configuration and processing of the prediction device will be described in detail with reference to Fig. 2. Fig. 2 is a block diagram showing an example of the configuration of the prediction device according to the embodiment.

[0025] 2 acquires observation results related to traffic volume at each point on a road shown on a map. For example, the observation results are a combination of the number of times an observation was performed at a certain point and the number of times a moving traffic entity was observed.

[0026] 2 is a functional block diagram showing the configuration of a prediction device according to an embodiment. As shown in FIG. 2, the prediction device 10 includes an interface unit 11, a storage unit 12, and a control unit 13.

[0027] The interface unit 11 is an interface for connecting the prediction device 10 to other devices. The prediction device 10 can exchange data with other devices via the interface unit 11.

[0028] For example, the prediction device 10 is a server that is connected to a road monitoring system and receives observation results from the system via an interface unit 11 .

[0029] The prediction device 10 then transmits the predicted traffic volume to a device for controlling traffic lights installed on the road, thereby enabling traffic lights to be controlled in accordance with the traffic volume.

[0030] For example, a device for controlling traffic lights can use the prediction results to shorten the time that traffic-based traffic spends waiting at traffic lights at locations with heavy traffic volume, and lengthen the time that traffic-based traffic spends waiting at traffic lights at locations with light traffic volume, thereby leveling out overall traffic volume.

[0031] The memory unit 12 and the control unit 13 of the prediction device 10 are realized by a computer or various circuits having, for example, a CPU (Central Processing Unit), a ROM (Read Only Memory), a RAM (Random Access Memory), a flash memory, an input / output port, etc.

[0032] The storage unit 12 corresponds to a RAM or a flash memory. The RAM or the flash memory can store model information 121, information on various programs, and the like.

[0033] The CPU of the computer also functions as the calculation unit 131, synthesis unit 132, prediction unit 133, and update unit 134 of the control unit 13 by reading and executing a program stored in, for example, the ROM.

[0034] The prediction device 10 may acquire the above-mentioned programs and various information via another computer or portable recording medium connected via a wired or wireless network.

[0035] Here, the model information 121 is parameters of a model for predicting traffic volume, etc. For example, when the model is expressed as a probability distribution (probability density function), the model information 121 is parameters for specifying the probability distribution.

[0036] The calculation unit 131 calculates the probability distribution of traffic volume for each point based on a Bayesian network that is constructed by regarding map data including nodes that are points and links that connect the nodes as a directed graph.

[0037] For example, the calculation unit 131 calculates a beta distribution as the probability distribution of a plurality of points. Note that the probability distribution calculated by the calculation unit 131 is not limited to the beta distribution, and may be a probability distribution other than the beta distribution.

[0038] Here, the calculation unit 131 calculates the probability distribution by Bayesian estimation. Therefore, the calculation result of the calculation unit 131 includes information on how likely the prediction is (certainty level, probability density).

[0039] As a result, the calculation unit 131 allows the AI ​​and analysts to decide how much they should trust the calculation results and make decisions and take action, making it possible to build a more flexible system.

[0040] The calculation results of the calculation unit 131 can be said to be probabilistic calculations of the traffic volume of the nodes by performing probability propagation from the measurement results. On the other hand, by propagating the probability distribution together with the confidence factor, the traffic volume between nodes can be predicted together with the confidence factor.

[0041] Therefore, the synthesis unit 132 obtains a synthetic distribution by approximately synthesizing the probability distributions of the multiple points calculated by the calculation unit 131.

[0042] For example, the synthesis unit 132 obtains a synthetic distribution using the geometric mean or harmonic mean of each parameter of the beta distribution, thereby obtaining a synthetic distribution that is easy to understand intuitively and has low computational cost.

[0043] Here, it is assumed that the calculation unit 131 calculates the probability distribution for the probability θ of node A and the probability φ of node B as shown in equations (1) and (2).

[0044]

number

[0045]

number

[0046] N, n, M, and m are all integers greater than or equal to 0 that represent the observation results. For example, N and n mean that, as a result of N observation attempts at node A, the movement of a traffic entity was observed n times. The same applies to M and m.

[0047] In other words, the more frequently a node is traveled by a transport entity, the larger the first term of the beta distribution parameter (e.g., n+α in equation (1)) and the smaller the second term (e.g., N-n+β in equation (1)).

[0048] Note that α and β are parameters that are determined by initial setting or learning, and are referred to as hyperparameters here.

[0049] The synthesis unit 132 approximately synthesizes the probability distributions as in equation (3) to obtain a synthetic distribution P(θ, φ).

[0050]

number

[0051] 3 is a diagram illustrating a method for synthesizing probability distributions according to an embodiment. Probability distribution 201 in Fig. 3 is the probability distribution of node A and corresponds to equation (1). Probability distribution 202 is the probability distribution of node B and corresponds to equation (1).

[0052] The probability distribution 201 and the probability distribution 202 may be expressed as a CPT (Conditional Probability Table).

[0053] Probability distribution 203 is a composite distribution and corresponds to equation (3). Here, the maximum likelihood estimation probability of probability distribution 201 is 0.3. The maximum likelihood estimation probability of probability distribution 202 is 0.6. The maximum likelihood estimation probability of probability distribution 203 is 0.6×0.3=0.18.

[0054] In this way, probability distribution 203 can be said to be a beta distribution whose maximum likelihood estimation probability is the product of the maximum likelihood estimation probability of node A and the maximum likelihood estimation probability of node B, and whose confidence and variance are intermediate between those of node A and node B.

[0055] Furthermore, probability distribution 203 is a probability distribution for predicting the traffic volume between a point corresponding to node A and a point corresponding to node B.

[0056] The prediction unit 133 predicts the traffic volume between points based on the probability distribution. In particular, the prediction unit 133 predicts the traffic volume between points based on the composite distribution.

[0057] Strictly combining probability distributions increases the number of dimensions, making it difficult to understand intuitively. On the other hand, approximating the combined distribution as in the embodiment makes it possible to provide intuitive traffic volume prediction results.

[0058] The update unit 134 updates the probability distribution by the weighted sum of the first parameter of the probability distribution calculated by the calculation unit 131 and the second parameter obtained from the observed value of the traffic volume.

[0059] The update unit 134 updates the model based on the difference between the prediction result of the model and the actual observation result, thereby coping with changes over time and quickly updating the model to the latest version with high responsiveness.

[0060] Here, the probability distribution of node A before updating, that is, the prior distribution, is assumed to be expressed by equation (4).

[0061]

number

[0062] On the other hand, the probability distribution based on the observation results of node A, that is, the posterior distribution, is expressed by equation (5).

[0063]

number

[0064] At this time, the update unit 134 updates the hyperparameters according to equation (6).

[0065]

number

[0066] a is the learning gain, and is the weight of the first term on the right side of equation (6). The update unit 134 calculates the learning gain a as shown in equation (8) based on the loss function Loss(θ) in equation (7).

[0067]

number

[0068]

number

[0069] In this way, the update unit 134 updates the probability distribution by the sum of the second parameter and a value obtained by multiplying the first parameter by a weight that is proportional to the absolute value of the error between the predicted value and the observed value according to the probability distribution and that decreases as the loss function that is inversely proportional to the ratio of the confidence of the probability distribution to the confidence of the observed value increases.

[0070] The first parameter corresponds to, for example, the value in brackets [ ] of the first term on the right side of equation (6), and the second parameter corresponds to, for example, the second term on the right side of equation (6).

[0071] This makes it possible to reduce the influence of learning (learning volatility) when the loss function is small, i.e., when the prediction error is small.

[0072] Fig. 4 is a diagram illustrating a method for updating a probability distribution according to an embodiment. As shown in Fig. 4, a prediction error (a gap between a model and an observation result) occurs between a prior distribution 301 and an observed distribution 401. Note that the observed distribution 401 is a probability distribution estimated from observed values.

[0073] The posterior distribution 302a is obtained by updating the prior distribution 301 without performing learning volatility (e.g., learning gain a=1). The posterior distribution 302b is obtained by updating the prior distribution 301 after performing learning volatility (e.g., learning gain a<1).

[0074] In this way, by performing learning evolution, the model can be updated so that the prediction results are closer to the actual observation results.

[0075] The update unit 134 may update the model when, for example, the learning gain a in equation (8) is equal to or greater than a threshold value (for example, 0.1).

[0076] The flow of the prediction process by the prediction device 10 will be described with reference to Fig. 5. Fig. 5 is a flowchart showing the flow of the prediction process according to the embodiment.

[0077] As shown in FIG. 5, first, the prediction device 10 calculates a probability distribution for each node by Bayesian estimation (step S101).

[0078] Next, the prediction device 10 approximately combines the probability distributions of the two nodes (step S102). For example, the prediction device 10 combines the probability distributions using equation (3).

[0079] Then, the prediction device 10 predicts the traffic volume between the nodes from the probability distribution obtained by the synthesis (step S103).

[0080] The flow of the update process performed by the prediction device 10 will be described with reference to Fig. 6. Fig. 6 is a flowchart showing the flow of the update process according to the embodiment.

[0081] 6, first, the prediction device 10 calculates the difference between the prior distribution (the model before updating) and the posterior distribution (the observation result) (step S201). For example, the difference is expressed by the loss function of equation (7).

[0082] Then, the prediction device 10 updates the model so that the influence of the prior distribution decreases as the difference increases (step S202). For example, the prediction device 10 updates the hyperparameters of the model as shown in equation (6) using the learning gain of equation (8).

[0083] As described above, the prediction device 10 according to the embodiment includes the calculation unit 131 and the prediction unit 133. The calculation unit 131 calculates a probability distribution of traffic volume for each point based on a Bayesian network configured by regarding map data including nodes that are points and links connecting the nodes as a directed graph. The prediction unit 133 predicts traffic volume between points based on the probability distribution.

[0084] In this way, the prediction device 10 predicts traffic volume using a method that uses a Bayesian network. As a result, according to the embodiment, traffic volume can be predicted using a method that has a clear mechanism and low calculation costs.

[0085] There are also known methods for predicting traffic volume using deep learning, etc. However, the mechanisms that lead to the results are not clear in such prediction methods, and only symptomatic measures based on the prediction results can be taken to address traffic problems, which may not lead to a fundamental solution.

[0086] In contrast, the prediction method of the embodiment has a clear mechanism for predicting traffic volume, and therefore can support fundamental solutions to traffic-related problems.

[0087] Further advantages and modifications will readily occur to those skilled in the art. Therefore, the invention in its broader aspects is not limited to the specific details and representative embodiments shown and described above. Accordingly, various modifications may be made without departing from the spirit or scope of the general inventive concept as defined by the appended claims and their equivalents. [Explanation of symbols]

[0088] 10 Prediction Device 11 Interface section 12 Storage section 13 Control Unit 121 Model Information 131 Calculation Department 132 Synthesis Section 133 Prediction Department 134 Update Department 201, 202, 203 Probability distribution 301 Prior distribution 302a, 302b Posterior distribution 401 Observation Distribution

Claims

1. a calculation unit that calculates a probability distribution of traffic volume for each point based on a Bayesian network that is constructed by regarding map data including nodes that are points and links connecting the nodes as a directed graph; a synthesis unit for obtaining a synthetic distribution by approximately synthesizing the probability distributions of the traffic volumes at the two points calculated by the calculation unit; a prediction unit that predicts traffic volume between the two points based on the composite distribution; A prediction device comprising:

2. the calculation unit calculates a beta distribution as a probability distribution of traffic volumes between the two points; The prediction device according to claim 1 , wherein the synthesis unit obtains a synthetic distribution using a geometric mean or a harmonic mean of each parameter of a beta distribution.

3. 3. The prediction device according to claim 1, further comprising an update unit that updates the probability distribution of traffic volume by a weighted sum of a first parameter of the probability distribution of traffic volume calculated by the calculation unit and a second parameter obtained from observed values ​​of traffic volume.

4. 4. The prediction device according to claim 3, wherein the update unit updates the probability distribution of traffic volume by the sum of the second parameter and a value obtained by multiplying the first parameter by a weight that is proportional to the absolute value of the error between the predicted value based on the probability distribution of traffic volume and the observed value, and that decreases as a loss function that is inversely proportional to a ratio of the certainty of the probability distribution of traffic volume to the certainty of the observed value increases.

5. A prediction method executed by a prediction device, comprising: a calculation step of calculating a probability distribution of traffic volume for each point based on a Bayesian network constructed by regarding map data including nodes that are points and links connecting the nodes as a directed graph; a combining step for obtaining a combined distribution by approximately combining the probability distributions of the traffic volumes at the two points calculated by the calculating step; a prediction step of predicting traffic volume between the two points based on the composite distribution; A prediction method comprising:

6. On the computer, a calculation step of calculating a probability distribution of traffic volume for each point based on a Bayesian network constructed by regarding map data including nodes that are points and links connecting the nodes as a directed graph; a combining step for obtaining a combined distribution by approximately combining the probability distributions of the traffic volumes at the two points calculated by the calculating step; a prediction step of predicting traffic volume between the two points based on the probability distribution; A prediction program characterized by executing the above.

Images