Machine learning program, method, and device

Abstract

To provide a program, method and device which build a machine learning model for high accuracy traffic simulation using a method expressing multiple traffic states as discriminable traffic demands.SOLUTION: In a machine learning device 10, a machine learning unit 20 includes: a first generation unit 22 which generates, based on route information indicating travel status of each of a plurality of mobile bodies in a specific geographical range at each of a plurality of points of time, traffic flow information indicating the number of mobile bodies located in route sections within the specific geographical range for each of the plurality of points of time; and a training unit 24 which trains a machine learning model for deriving a degree of traffic congestion corresponding to the traffic flow information, using training data that includes the generated traffic flow information as input feature values and includes information indicating a degree of traffic congestion in the specific geographical range at a point of time corresponding to the traffic flow information as label information.SELECTED DRAWING: Figure 6

Description

[Technical field]

[0001] The disclosed technology relates to a machine learning program, a machine learning method, and a machine learning device. [Background technology]

[0002] Conventionally, there is a technology that uses traffic simulation to optimize public transportation schedules. For example, when determining the operation schedule of a special bus during an event, it is necessary to set the departure and arrival times of the special bus according to the traffic conditions to avoid the occurrence and expansion of congestion. In such a case, a traffic simulation is used to search for an optimal schedule that does not cause congestion.

[0003] Building a traffic simulation device is very time-consuming. In addition, the computational load during simulation is high. However, by extracting traffic demand and traffic density from probe data and learning the relationship between the two using machine learning models such as neural networks, it is possible to build a highly accurate alternative simulation model (surrogate model). Because the computational load of the surrogate model is low, it can expand the range of applications of task optimization through simulation. For example, it can be applied to optimizing daily schedules.

[0004] As a technology related to traffic simulation using a surrogate model, a machine learning device that efficiently estimates pedestrian and traffic flows has been proposed. The device acquires a first parameter that represents an environment and a second parameter that represents an attribute of the movement of each of a plurality of moving objects in the environment, and classifies the plurality of moving objects into a plurality of groups based on the second parameter. The device also generates a third parameter that indicates the number of moving objects classified into each of the plurality of groups, inputs the first parameter and the third parameter into a machine learning model, and generates estimated information regarding the movement of the plurality of moving objects in the environment. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] JP 2022-131393 A Summary of the Invention [Problem to be solved by the invention]

[0006] When training a surrogate model for traffic simulation, training data is generally used in which the traffic demand expressed in an Origin-Destination (OD) table is used as the feature and the traffic density is used as the correct label. However, there are cases where the feature expressed in an OD table cannot properly distinguish between various traffic conditions, and training of the surrogate model does not proceed well.

[0007] In one aspect, the disclosed technology aims to build a highly accurate machine learning model for traffic simulation using a method of expressing traffic demand that can distinguish between various traffic conditions. [Means for solving the problem]

[0008] In one aspect, the disclosed technology generates traffic flow information indicating the number of moving objects located in each route section within a predetermined geographical range for each of the multiple time points based on route information indicating the movement status of each of multiple moving objects in the predetermined geographical range for the multiple time points.The disclosed technology then trains a machine learning model that derives the traffic congestion level according to the traffic flow information using training data in which the traffic flow information is used as input features and information indicating the traffic congestion level in the predetermined geographical range at the time point corresponding to the traffic flow information is used as label information. Effect of the Invention

[0009] One aspect of this invention is that it has the effect of enabling the construction of a highly accurate machine learning model for traffic simulation using a method of expressing traffic demand that can distinguish between various traffic conditions. [Brief description of the drawings]

[0010] [Figure 1] FIG. 13 is a diagram for explaining a traffic simulation. [Diagram 2] FIG. 1 is a schematic diagram showing an example of a road network. [Diagram 3] FIG. 1 is a diagram for explaining the machine learning phase of a surrogate model. [Figure 4] FIG. 13 is a diagram for explaining an optimization phase using a surrogate model. [Diagram 5] FIG. 13 is a diagram for explaining a problem when using features of an OD table representation. [Figure 6] FIG. 1 is a functional block diagram of a machine learning device according to an embodiment of the present invention. [Figure 7] FIG. 2 is a diagram for explaining generation of traffic flow information. [Figure 8] FIG. 13 is a diagram for explaining conversion of probe data into a link sequence. [Figure 9] FIG. 11 is a diagram showing an example of traffic flow information in which link sequences are aggregated by time. [Figure 10] FIG. 13 is a diagram showing an example of traffic flow information in which link sequences are aggregated by departure time. [Figure 11] FIG. 13 is a diagram for explaining feature amounts in the case of OD table representation. [Figure 12] FIG. 4 is a diagram for explaining feature amounts according to the present embodiment. [Figure 13] FIG. 13 is a diagram for explaining generation of route information based on an OD forecast value and supplementary information. [Figure 14] FIG. 1 is a block diagram showing a schematic configuration of a computer that functions as a machine learning device. [Figure 15] 1 is a flowchart illustrating an example of a machine learning process. [Figure 16] 13 is a flowchart illustrating an example of an optimization process. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] Hereinafter, an example of an embodiment of the disclosed technology will be described with reference to the drawings.

[0012] <Overview and Issues of Traffic Simulation and Surrogate Model> Before describing the details of the embodiment, an overview and problems of general traffic simulation and surrogate models will be described.

[0013] As shown in FIG. 1, traffic simulation is a method of inputting traffic demand, simulating traffic conditions according to the traffic demand, and outputting the simulation. For example, as shown in FIG. 2, traffic simulation is performed using a road network represented by a plurality of nodes and links connecting the nodes. In the example of FIG. 2, black circles represent nodes, and connecting lines connecting the nodes represent links. In addition, "ni (i=1,2,...)" written next to a node is the identification number of the node, and "ej (j=1,2,...)" written next to a link is the identification number of the link. In the following, a node with identification number ni is represented as "node ni", and a link with identification number ej is represented as "link ej". In addition, a link may be represented by connecting the node numbers of the nodes at both ends of the link. For example, link e1 is "n1-n2". In addition, a link is an example of a "route section" in the disclosed technology.

[0014] Traffic demand is represented, for example, by an OD table showing the number of moving objects (people, vehicles, etc.) moving between each point (node). The OD table is represented as a matrix of origin nodes and destination nodes, with the number of moving objects moving from the corresponding origin to the destination stored in each square of the matrix. In addition, a three-dimensional tensor as shown in FIG. 1 may be used as traffic demand using OD by time period. In addition, traffic conditions are, for example, traffic density represented by traffic volume for each link.

[0015] A machine learning model (surrogate model) for traffic simulation is a lightweight alternative model that reproduces the input / output relationship of a traffic simulation as shown in FIG. 1. As shown in FIG. 3, in the machine learning phase of the machine learning model, an information processing device extracts traffic demand and road networks, which are feature quantities, from geographic data and multiple probe data, and also extracts traffic density, which is a correct label. Probe data is data in which location information (latitude, longitude, etc.) of each moving object is recorded together with a timestamp. Probe data is acquired by smartphones, various sensors, etc., and collected via a network. Geographic data is data in which various geographic information, including road network information, is recorded.

[0016] The information processing device then inputs the extracted traffic demand and road network as features into a machine learning model composed of a neural network or the like, and acquires the traffic density inferred by the machine learning model. The information processing device then trains the machine learning model by updating parameters of the machine learning model so as to minimize the error between the acquired traffic density and the traffic density that is the correct label.

[0017] In the optimization phase using the trained machine learning model, as shown in FIG. 4, the information processing device inputs the road network and traffic demand into the machine learning model to obtain the traffic density, which is the objective function. FIG. 4 shows an example of optimizing a bus operation schedule when an event is held. In the example of FIG. 4, the information processing device inputs the event traffic demand for each bus operation schedule proposal as a decision variable into the machine learning model, with the road network and daily traffic demand forecast values ​​as constants. The road network may also be treated as a decision variable. Then, the information processing device compares the traffic density obtained for each proposal, and selects and outputs the proposal with the minimum traffic density (congestion) as the optimal proposal. The optimization may be performed using an existing method such as a genetic algorithm.

[0018] When features expressed in an OD table as described above are used as features for training a machine learning model (surrogate model) for traffic simulation, situations may arise in which various traffic conditions cannot be properly distinguished, and training of the machine learning model may not proceed smoothly.

[0019] For example, consider situations 1 and 2 as shown in FIG. 5. In situation 1, there are 100 pieces of probe data showing route 1 (dashed arrow in FIG. 5) of nodes n1 → n2 → n3 → n4 → n8 → n9, and 50 pieces of probe data showing route 2 (dash-dotted arrow in FIG. 5) of nodes n1 → n2 → n3 → n4 → n5 → n7. In situation 2, there are 100 pieces of probe data showing route 1 similar to situation 1, and 50 pieces of probe data showing route 3 (dash-dotted arrow in FIG. 5) of nodes n1 → n2 → n6 → n5 → n7. In this way, even though situations 1 and 2 are different, when these situations are expressed in OD tables, as shown in the lower diagram of FIG. 5, they become the same OD table. In other words, when a machine learning model is trained using this OD table as a feature, it is not possible to distinguish between the two situations, and it is not possible to build a highly accurate machine learning model.

[0020] Therefore, it is conceivable to distinguish more situations by using route information. However, if the route information of each moving body is used as a feature directly, the number of dimensions of the feature will be too large. The route information of each moving body is sequence data of nodes that indicate the moving path of the moving body, such as the above-mentioned n1 → n2 →.... The route information is not limited to sequence data of nodes, but may be sequence data of links, or sequence data of position information (latitude and longitude). If the route information is used as a feature directly, the number of dimensions of the feature will be an enormous number of dimensions, for example, the number of nodes included in the route information (hundreds) × the number of moving bodies (tens of thousands to millions).

[0021] Therefore, in this embodiment, a traffic demand is characterized by an expression method capable of distinguishing various traffic conditions, using information summarizing route information for each mobile body, specifically information on the number of mobile bodies passing through each link (link traffic flow). In this case, the number of dimensions of the feature is, for example, about the number of links (several thousand) x time intervals (several dozens). In this embodiment, a highly accurate machine learning model for traffic simulation is constructed using a traffic demand expression method capable of distinguishing various traffic conditions.

[0022] <Machine Learning Device According to the Present Embodiment> 6, the machine learning device 10 according to this embodiment includes a machine learning unit 20 and an inference unit 40. A machine learning model 30, which is a surrogate model for traffic simulation, is stored in a predetermined storage area of ​​the machine learning device 10. The machine learning model 30 is configured, for example, by a neural network.

[0023] The machine learning unit 20 is a functional unit that functions in the machine learning phase. In the machine learning phase, a plurality of probe data and geographic data are input to the machine learning device 10. The machine learning unit 20 includes a first generation unit 22 and a training unit 24.

[0024] The first generation unit 22 acquires a plurality of probe data and geographic data input to the machine learning device 10. The first generation unit 22 generates traffic flow information indicating the number of mobile objects located on each link within a predetermined geographic range for each of the plurality of time points, based on route information indicating the movement status of each of a plurality of mobile objects in a predetermined geographic range for each of the plurality of time points. The predetermined geographic range is an area indicated by a road network included in the geographic data.

[0025] Specifically, the first generating unit 22 generates traffic flow information by tallying up the traffic flow of each link of the road network based on the route information indicated by each of the multiple probe data, as shown in Fig. 7. This makes it possible to generate features that distinguish between situations, such as situation 1 and situation 2 shown in Fig. 5, even in situations where the features are the same in the OD table representation.

[0026] More specifically, the first generating unit 22 shapes the probe data. For example, as shown in the upper diagram of Fig. 8, when the measurement interval (interval of the "time" item) of the probe data expressed in the data format of {ID (identification information of a moving object), time, latitude, longitude} is not constant, the first generating unit 22 may perform interpolation so that the time interval becomes constant. For example, the first generating unit 22 may perform interpolation so that the probe data becomes data at one-minute intervals.

[0027] The first generating unit 22 associates the shaped probe data with the road network and compares the position information (latitude and longitude) of the mobile object with the position information of each link. Then, as shown in the lower diagram of FIG. 8, the first generating unit 22 converts the probe data into sequence data of links where the mobile object exists at each time (hereinafter, referred to as "link sequence"). The first generating unit 22 generates traffic flow information by aggregating the number of mobile objects existing on each link at each time based on the link sequence of each mobile object converted from the probe data. FIG. 9 shows traffic flow information aggregating the link sequences shown in the lower diagram of FIG. 8.

[0028] Here, the probe data includes, for example, the effects of traffic signal delays, congestion delays, and the like. It is desirable that the features used to train the machine learning model 30 exclude these effects. This is because, at the time of inference, it is impossible to provide traffic demand including the effects of traffic signal delays, congestion delays, and the like, i.e., the effects of interactions between moving objects or interactions with a road network. In the first place, the purpose of traffic simulation is to simulate traffic conditions taking into account the effects of delays, etc., from traffic demand equivalent to a route planned in advance, excluding the effects of these delays, etc. Therefore, as described above, it is desirable that the features used to train the machine learning model 30 do not include the effects of delays, etc.

[0029] Therefore, the first generating unit 22 may create traffic flow information from which information on the time during movement is deleted. Specifically, the first generating unit 22 associates the link series of each moving object with the time corresponding to the starting point of the link series, as shown in the upper diagram of Fig. 10. Then, the first generating unit 22 generates traffic flow information that is aggregated by associating the link series with the time corresponding to the starting point, as shown in the lower diagram of Fig. 10. In this way, traffic flow information equivalent to a route planned in advance is generated, excluding the effects of delays due to traffic signals, delays due to congestion, and the like.

[0030] In this way, the traffic flow information, which is the feature generated by the first generating unit 22, is the feature of the links constituting the road network. For example, in the case of the feature of the OD table representation, as shown in FIG. 11, each node holds information on the number of route information starting from that node and ending at each of the other nodes. That is, each node holds information for the number of nodes included in the road network minus 1. On the other hand, the feature of this embodiment is such that each link holds one piece of traffic flow information for that link, as shown in FIG. 12.

[0031] The first generating unit 22 also generates information indicating the degree of traffic congestion in a predetermined geographical area at a time point corresponding to the traffic flow information based on the probe data as a correct label of traffic density. For example, the first generating unit 22 generates a link traffic flow represented by the number of mobile objects existing on each link at each time point as a correct label of traffic density.

[0032] The training unit 24 trains the machine learning model 30, which derives the degree of traffic congestion according to the traffic flow information, using training data including the input features and a correct answer label of the traffic density, with the traffic flow information generated by the first generation unit 22 as input features. Specifically, the training unit 24 inputs the generated traffic flow information as features to the machine learning model 30, and acquires the traffic density inferred by the machine learning model 30. Then, the training unit 24 trains the machine learning model 30 by updating parameters of the machine learning model 30 so as to minimize an error between the acquired traffic density and the traffic density generated as the correct answer label.

[0033] The inference unit 40 is a functional unit that functions in the optimization phase. In the optimization phase, supplementary information (details will be described later) that is at least one of domain knowledge related to traffic demand, traffic flow in a partial route section, and movement status of a mobile object is input to the machine learning device 10. The inference unit 40 includes a preprocessing unit 42, a second generation unit 44, and an optimization unit 46.

[0034] Here, in this embodiment, the feature quantity input to the machine learning model 30 is traffic flow information as shown in FIG. 9 or the lower diagram of FIG. 10. This traffic flow information is generated by aggregating route information of the mobile object. However, in the optimization phase, probe data like that in the machine learning phase is not obtained, and in many cases, only OD forecast values ​​are obtained as traffic demand. In some cases, a partial link traffic flow prediction value may be given. Furthermore, as a basic principle for route selection, domain knowledge regarding the movement status of the mobile object, such as road distance information, OD type information, toll information, and information on the driver's behavior tendency, may be available. Such partial link traffic flow prediction value and domain knowledge are collectively referred to as "supplementary information".

[0035] Therefore, when the traffic demand, i.e., the OD forecast value, for each of a plurality of combinations of departure points and arrival points is given, the preprocessing unit 42 generates route information by allocating each of the plurality of combinations to a plurality of routes. Also, when supplementary information is given in addition to the OD forecast value, the preprocessing unit 42 allocates each of the plurality of combinations to a plurality of routes based on the supplementary information.

[0036] For example, as shown in the upper diagram of FIG. 13, the OD forecast value of a route with node n1 as the starting point and node n7 as the destination point and the link traffic flow of links n2-n3 and n2-n6 are given. Furthermore, as shown in the lower diagram of FIG. 13, there are route 1 (arrow with dashed lines) of n1→n2→n3→n4→n5→n7 and route 2 (arrow with dashed lines) of n1→n2→n6→n5→n7 as routes that satisfy this OD. In this case, the preprocessing unit 42 distributes the OD forecast value to route 1 and route 2 based on the ratio of the link traffic flow between the links where the two routes branch off. In the example of FIG. 13, the preprocessing unit 42 distributes the OD forecast value of 100 from n1→n7 to route 1:route 2=20:80 based on the ratio (5:20) of the link traffic flow between link n2-n3 and link n2-n6. In this manner, the pre-processing unit 42 generates route information from the OD forecast values ​​and partial link traffic flow prediction values.

[0037] In addition, when a partial link traffic flow prediction value is not available, route information may be generated by distributing OD forecast values ​​evenly or randomly to routes that satisfy OD. Also, the distribution may be performed using domain knowledge. For example, based on the domain knowledge that "people choose routes that minimize travel time," route information may be generated by distributing the OD forecast value to routes that minimize travel time among routes that satisfy OD. Also, for example, if the OD forecast value is for a late-night time period, the OD may be regarded as the OD of a transport truck, and the domain knowledge that "transport trucks tend to prioritize time savings even if it costs a little more." may be used. Also, if the OD forecast value is for a daytime time period, the OD may be regarded as the OD of a passenger car, and the domain knowledge that "passenger cars tend to prioritize low fees over time savings." may be used.

[0038] When multiple domain knowledge are available, the allocation to each route may be determined by appropriately combining these domain knowledge. The more domain knowledge that is available, the more realistic route information can be generated from the given OD forecast value. In other words, the more domain knowledge that is available, the more distinguishable traffic conditions can be, and the more accurate the inference by the machine learning model 30 can be.

[0039] The second generation unit 44, like the first generation unit 22, uses the route information generated by the pre-processing unit 42 to generate link sequences from the route information, aggregates the link sequences by time or by departure time, and generates traffic flow information that serves as features to be input to the machine learning model 30.

[0040] The optimization unit 46 predicts traffic density by inputting the traffic flow information generated by the second generation unit 44 into the machine learning model 30. The optimization unit 46 predicts traffic density for each traffic demand according to a proposal such as a schedule to be simulated, and selects and outputs the proposal with the smallest traffic density as the optimal proposal.

[0041] The machine learning device 10 may be realized by, for example, a computer 50 shown in FIG. 14. The computer 50 includes a CPU (Central Processing Unit) 51, a GPU (Graphics Processing Unit) 52, a memory 53 as a temporary storage area, and a non-volatile storage device 54. The computer 50 also includes an input / output device 55 such as an input device and a display device, and an R / W (Read / Write) device 56 that controls reading and writing of data to a storage medium 59. The computer 50 also includes a communication I / F (Interface) 57 that is connected to a network such as the Internet. The CPU 51, the GPU 52, the memory 53, the storage device 54, the input / output device 55, the R / W device 56, and the communication I / F 57 are connected to each other via a bus 58.

[0042] The storage device 54 is, for example, a hard disk drive (HDD), a solid state drive (SSD), a flash memory, or the like. The storage device 54 as a storage medium stores a machine learning program 60 for causing the computer 50 to function as the machine learning unit 20, and an inference program 70 for causing the computer 50 to function as the inference unit 40. The machine learning program 60 has a first generation process control instruction 62 and a training process control instruction 64. The inference program 70 has a pre-processing process control instruction 72, a second generation process control instruction 74, and an optimization process control instruction 76. The storage device 54 also has an information storage area 80 in which information constituting the machine learning model 30 is stored.

[0043] The CPU 51 reads the machine learning program 60 from the storage device 54, loads it in the memory 53, and sequentially executes the control instructions of the machine learning program 60. The CPU 51 operates as the first generation unit 22 shown in FIG. 6 by executing the first generation process control instruction 62. The CPU 51 also operates as the training unit 24 shown in FIG. 6 by executing the training process control instruction 64. The CPU 51 also reads information from the information storage area 80, and loads the machine learning model 30 in the memory 53. As a result, the computer 50 that has executed the machine learning program 60 functions as the machine learning unit 20 of the machine learning device 10.

[0044] The CPU 51 reads out the inference program 70 from the storage device 54, expands it in the memory 53, and sequentially executes the control instructions of the inference program 70. The CPU 51 operates as the pre-processing unit 42 shown in FIG. 6 by executing the pre-processing process control instruction 72. The CPU 51 also operates as the second generation unit 44 shown in FIG. 6 by executing the second generation process control instruction 74. The CPU 51 also operates as the optimization unit 46 shown in FIG. 6 by executing the optimization process control instruction 76. The CPU 51 also reads out information from the information storage area 80, and expands the machine learning model 30 in the memory 53. As a result, the computer 50 that has executed the inference program 70 functions as the inference unit 40 of the machine learning device 10.

[0045] The CPU 51 that executes the program is hardware. In addition, part of the program may be executed by the GPU 52.

[0046] In addition, the functions realized by the machine learning program 60 and the inference program 70 may be realized, for example, by a semiconductor integrated circuit, more specifically, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), etc.

[0047] Next, the operation of the machine learning device 10 according to the present embodiment will be described. In the machine learning phase, when a plurality of probe data and geographic data are input to the machine learning device 10, the machine learning device 10 executes the machine learning process shown in FIG. 15. In addition, in the optimization phase, when traffic demand and supplementary information are input to the machine learning device 10, the machine learning device 10 executes the optimization process shown in FIG. 16. Note that the machine learning process is an example of a machine learning method of the disclosed technology. Below, the machine learning process and the optimization process will each be described in detail.

[0048] First, the machine learning process shown in FIG. 15 will be described.

[0049] In step S10, the first generation unit 22 acquires a plurality of pieces of probe data and geographic data input to the machine learning device 10. Next, in step S12, the first generation unit 22 associates the probe data with a road network and converts the probe data into a link sequence.

[0050] Next, in step S14, the first generation unit 22 aggregates all links included in the link series of each mobile object by the time corresponding to the starting point of the link series, and generates traffic flow information by counting the mobile objects present on each link at each aggregated time. Next, in step S16, the first generation unit 22 generates link traffic flows represented by the number of mobile objects present on each link at each time from the probe data as correct answer labels of traffic density.

[0051] Next, in step S18, the training unit 24 uses the traffic flow information generated in step S14 as features and the correct label generated in step S16 as training data to train the machine learning model 30. Then, the training unit 24 stores the trained machine learning model 30 in a predetermined storage area, and the machine learning process ends.

[0052] Next, the optimization process shown in FIG. 16 will be described.

[0053] A loop process in step S20 is executed for each plan of a schedule or the like for which a traffic simulation is to be performed. Specifically, in step S22, the pre-processing unit 42 acquires the OD forecast value, which is the traffic demand, and the supplementary information input to the machine learning device 10. Next, in step S24, the pre-processing unit 42 generates route information by allocating the OD forecast value to each route corresponding to the OD based on the supplementary information.

[0054] Next, in step S26, the second generation unit 44 converts the route information into a link sequence. Next, in step S28, the second generation unit 44 aggregates all links included in the link sequence of each moving object by the time corresponding to the starting point of the link sequence, and counts the moving objects present on each link at each aggregated time, thereby generating traffic flow information.

[0055] Next, in step S30, the optimization unit 46 predicts traffic density by inputting the traffic flow information generated in step S28 above into the machine learning model 30. When the processing of steps S22 to S30 above is completed for all plans, the loop processing of step S20 is terminated and the process proceeds to step S32.

[0056] In step S32, the optimization unit 46 selects and outputs the plan with the smallest predicted traffic density as the optimal plan from among the plans, and the optimization process ends.

[0057] As described above, the machine learning device according to the present embodiment generates traffic flow information indicating the number of mobile objects located on each link in a road network at each of multiple time points based on the probe data of each of multiple mobile objects in a predetermined geographical area. The machine learning device then trains a machine learning model that derives traffic density according to the traffic flow information using training data in which the traffic flow information is used as input features and the traffic density at the time point corresponding to the traffic flow information is used as label information. This makes it possible to build a highly accurate machine learning model for traffic simulation using a method of expressing traffic demand that can distinguish between various traffic conditions.

[0058] In the above machine learning process and optimization process, we have described the case where traffic flow information generated from link sequences aggregated by departure time is used as a feature. However, traffic flow information in which link sequences are aggregated by time, as shown in Figure 9, may also be used as a feature.

[0059] In addition, in the above embodiment, the machine learning unit and the inference unit are configured in the same computer, but they may be configured in separate computers.

[0060] In the above embodiment, the machine learning program and the inference program are stored (installed) in advance in the storage device, but this is not limiting. The program according to the disclosed technology may be provided in a form stored in a storage medium such as a CD-ROM, a DVD-ROM, or a USB memory.

[0061] Regarding the above embodiment, the following supplementary notes are further disclosed.

[0062] (Appendix 1) generating traffic flow information indicating the number of mobile objects located in each route section within the predetermined geographical range for each of the plurality of times based on route information indicating the movement status of each of the plurality of mobile objects within the predetermined geographical range for each of the plurality of times; training a machine learning model that derives a traffic congestion level according to the traffic flow information, using training data in which the traffic flow information is used as an input feature and information indicating a traffic congestion level in the specified geographical area at a time point corresponding to the traffic flow information is used as label information; A machine learning program that causes a computer to perform processes including:

[0063] (Appendix 2) The machine learning program of Appendix 1, wherein generating the traffic flow information includes converting the position information of the moving body at each time point included in the route information into position information at a time point corresponding to a starting point included in the route information, and counting, for each route section, the number of route information that includes position information corresponding to the same route section based on one or more route information that have the same time point corresponding to the starting point.

[0064] (Appendix 3) 3. The machine learning program of claim 1 or 2, for causing the computer to execute a process further including inputting traffic flow information to be inferred into a trained machine learning model and inferring traffic congestion based on the traffic flow information to be inferred.

[0065] (Appendix 4) 4. The machine learning program of claim 3, wherein, when traffic demand for each of a plurality of combinations of departure points and arrival points is given, traffic flow information generated based on route information generated by allocating each of the plurality of combinations to a plurality of routes is input to the trained machine learning model as the traffic flow information to be inferred.

[0066] (Appendix 5) The machine learning program of claim 4, which allocates each of the plurality of combinations to the plurality of routes based on supplementary information which is at least one of domain knowledge regarding traffic flow in partial route sections and movement conditions of mobile objects.

[0067] (Appendix 6) generating traffic flow information indicating the number of mobile objects located in each route section within the predetermined geographical range for each of the plurality of times based on route information indicating the movement status of each of the plurality of mobile objects within the predetermined geographical range for each of the plurality of times; training a machine learning model that derives a traffic congestion level according to the traffic flow information, using training data in which the traffic flow information is used as an input feature and information indicating a traffic congestion level in the specified geographical area at a time point corresponding to the traffic flow information is used as label information; A machine learning method in which a computer performs a process including:

[0068] (Appendix 7) The machine learning method of claim 6, wherein generating the traffic flow information includes converting the position information of the moving body at each time point included in the route information into position information at a time point corresponding to a starting point included in the route information, and counting, for each route section, the number of route information that includes position information corresponding to the same route section based on one or more route information having the same time point corresponding to the starting point.

[0069] (Appendix 8) 8. The machine learning method according to claim 6 or 7, wherein the computer executes a process further including inputting traffic flow information to be inferred into a trained machine learning model, and inferring traffic congestion according to the traffic flow information to be inferred.

[0070] (Appendix 9) 9. The machine learning method according to claim 8, in which, when traffic demand for each of a plurality of combinations of departure points and arrival points is given, traffic flow information generated based on route information generated by allocating each of the plurality of combinations to a plurality of routes is input to the trained machine learning model as the traffic flow information to be inferred.

[0071] (Appendix 10) 10. The machine learning method of claim 9, further comprising allocating each of the combinations to the routes based on supplementary information which is at least one of domain knowledge regarding traffic flow in partial route sections and movement conditions of mobile objects.

[0072] (Appendix 11) a first generation unit that generates traffic flow information indicating the number of moving objects located in each route section within a predetermined geographical range for each of the plurality of time points based on route information indicating the movement status of each of the plurality of moving objects within the predetermined geographical range for each of the plurality of time points; a training unit that trains a machine learning model that derives a traffic congestion level according to the traffic flow information by using training data in which the traffic flow information is an input feature and information indicating a traffic congestion level in the specified geographical area at a time point corresponding to the traffic flow information is label information; and A machine learning device including:

[0073] (Appendix 12) The machine learning device of claim 11, wherein the first generation unit regards position information of the moving body at each time point included in the route information as position information at a time point corresponding to a starting point included in the route information, and counts the number of route information that include position information corresponding to the same route section for each route section based on one or more of the route information having the same time point corresponding to the starting point.

[0074] (Appendix 13) 13. The machine learning device according to claim 11 or 12, further comprising an inference unit that inputs traffic flow information to be inferred into a trained machine learning model and infers traffic congestion according to the traffic flow information to be inferred.

[0075] (Appendix 14) The machine learning device according to claim 13, wherein the inference unit, when traffic demand for each of a plurality of combinations of departure points and arrival points is given, inputs traffic flow information generated based on route information generated by allocating each of the plurality of combinations to a plurality of routes to the trained machine learning model as traffic flow information to be inferred.

[0076] (Appendix 15) The machine learning device of claim 14, wherein the inference unit allocates each of the combinations to the multiple routes based on supplementary information which is at least one of domain knowledge regarding traffic flow in partial route sections and movement conditions of moving objects.

[0077] (Appendix 16) generating traffic flow information indicating the number of mobile objects located in each route section within the predetermined geographical range for each of the plurality of times based on route information indicating the movement status of each of the plurality of mobile objects within the predetermined geographical range for each of the plurality of times; training a machine learning model that derives a traffic congestion level according to the traffic flow information, using training data in which the traffic flow information is used as an input feature and information indicating a traffic congestion level in the specified geographical area at a time point corresponding to the traffic flow information is used as label information; A non-transitory storage medium storing a machine learning program for causing a computer to execute processes including the above. [Explanation of symbols]

[0078] 10 Machine Learning Device 20 Machine Learning Department 22 1st generation part 24 Training Department 30 Machine Learning Models 40 Reasoning part 42 Pretreatment section 44 Second generation part 46 Optimization Department 50 Computer 51 CPU 52 GPU 53 Memory 54 Storage device 55 Input / Output Devices 56 R / W device 57 Communication I / F 58 Bus 59 Storage medium 60 Machine Learning Programs 62 First generation process control instruction 64 Training process control instructions 70 Inference Programs 72 Pretreatment process control instructions 74 Second generation process control instruction 76 Optimization Process Control Instructions 80 Information storage area

Claims

1. generating traffic flow information indicating the number of mobile objects located in each route section within the predetermined geographical range for each of the plurality of times based on route information indicating the movement status of each of the plurality of mobile objects within the predetermined geographical range for each of the plurality of times; training a machine learning model that derives a traffic congestion level according to the traffic flow information, using training data in which the traffic flow information is used as an input feature and information indicating a traffic congestion level in the specified geographical area at a time point corresponding to the traffic flow information is used as label information; A machine learning program that causes a computer to perform processes including:

2. The machine learning program of claim 1, wherein generating the traffic flow information includes converting the position information of the moving body at each time point included in the route information into position information at a time point corresponding to a starting point included in the route information, and counting, for each route section, the number of route information including position information corresponding to the same route section based on one or more route information having the same time point corresponding to the starting point.

3. 3. The machine learning program of claim 1 or 2, for causing the computer to execute a process further including inputting traffic flow information to be inferred into a trained machine learning model and inferring traffic congestion based on the traffic flow information to be inferred.

4. The machine learning program of claim 3, wherein when traffic demand for each of a plurality of combinations of departure points and arrival points is given, traffic flow information generated based on route information generated by allocating each of the plurality of combinations to a plurality of routes is input to the trained machine learning model as the traffic flow information to be inferred.

5. The machine learning program according to claim 4 , further comprising: allocating each of the plurality of combinations to the plurality of routes based on supplementary information which is at least one of domain knowledge regarding traffic flow in partial route sections and movement conditions of mobile objects.

6. generating traffic flow information indicating the number of mobile objects located in each route section within the predetermined geographical range for each of the plurality of times based on route information indicating the movement status of each of the plurality of mobile objects within the predetermined geographical range for each of the plurality of times; training a machine learning model that derives a traffic congestion level according to the traffic flow information, using training data in which the traffic flow information is used as an input feature and information indicating a traffic congestion level in the specified geographical area at a time point corresponding to the traffic flow information is used as label information; A machine learning method in which a computer performs a process including:

7. a first generation unit that generates traffic flow information indicating the number of moving objects located in each route section within a predetermined geographical range for each of a plurality of time points based on route information indicating the movement status of each of a plurality of moving objects within the predetermined geographical range for the plurality of time points; a training unit that trains a machine learning model that derives a traffic congestion level according to the traffic flow information by using training data in which the traffic flow information is an input feature and information indicating a traffic congestion level in the specified geographical area at a time point corresponding to the traffic flow information is label information; and A machine learning device including:

Images