Information processing device, information processing method, and program for estimating radio wave propagation characteristics

Abstract

To estimate propagation characteristics of a radio wave with high accuracy by constructing a model having high generation performance.SOLUTION: An information processing device uses image data showing heights of buildings or occupancy of the buildings in each of a plurality of areas in a prescribed size included in a prescribed geographical range as one pixel as input data, acquires a first learned model obtained by machine learning using information of a propagation path of a radio wave from a transmission point to a reception point based on the arrangement of a building within the prescribed geographical range as teacher data, uses an estimation value of the propagation path obtained by inputting the image data in the first learned model as an input, acquires a second learned model by machine learning using a measured value of the propagation characteristics of a radio wave at the reception point as teacher data, connects the first learned model and the second learned model, and generates a model for estimating propagation characteristics of the radio wave from the image data.SELECTED DRAWING: Figure 8

Description

[Technical Field]

[0001] The present invention relates to a technology for estimating radio wave propagation characteristics using machine learning. [Background technology]

[0002] In wireless communication systems, it is important that radio waves reach the receiving end with sufficient power. Cellular communication systems address dead zones where radio waves from base station devices do not reach with sufficient power. Conventionally, to address such dead zones, measuring instruments were used to measure radio waves actually transmitted from base station devices, which was extremely time-consuming. Ray tracing is another technology for estimating radio wave propagation characteristics. Ray tracing estimates propagation characteristics by calculating how a ray representing a radio wave is emitted from a transmitting point in a modeled environment, such as a building, and how the ray reaches a receiving point after passing through reflections and diffractions. However, this calculation requires a huge amount of computation, making it difficult to obtain results quickly. In response to this issue, Patent Document 1 proposes a technology for estimating propagation characteristics by applying a convolutional neural network (CNN) to map data containing receiving points to extract urban structure parameters as features, and then inputting the extracted features into a fully connected neural network (FNN). This makes it possible to use map data to quickly estimate propagation characteristics without actually measuring the radio wave environment and with a small amount of calculation. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2019-122008 Summary of the Invention [Problem to be solved by the invention]

[0004] In the technology described in Patent Document 1, when extracting features from map data, image similarity is learned, which may result in the extraction of physically meaningless features. In such cases, while highly accurate estimation of radio wave propagation characteristics is possible in areas where training data used in machine learning exists, the estimation accuracy of radio wave propagation characteristics may deteriorate in areas where training data does not exist. In other words, a model with poor generalization performance may be generated. [Means for solving the problem]

[0005] The present invention provides a technology that enables a model with high generalization performance to be constructed and radio wave propagation characteristics to be estimated with high accuracy.

[0006] An information processing device according to one embodiment of the present invention has a first acquisition means that acquires a first trained model through machine learning using image data that shows, as input data, the building height or building occupancy rate in each of multiple areas of a predetermined size included in a predetermined geographical range, and information on the propagation path of radio waves from a transmission point to a reception point based on the arrangement of buildings in the predetermined geographical range as training data; a second acquisition means that acquires a second trained model through machine learning using, as input data, an estimated value of the propagation path obtained by inputting the image data into the first trained model, and using, as training data, actual measured values ​​of the propagation characteristics of radio waves at the reception point; and a generation means that link the first trained model and the second trained model to generate a model that estimates the propagation characteristics of radio waves from the image data.

[0007] An information processing device according to another aspect of the present invention has a first acquisition means that inputs first image data corresponding to a second geographical range, the first image data showing the building height or building occupancy rate as one pixel in each of a plurality of areas of a predetermined size included in a first geographical range, and inputs second image data of the same format as the input data corresponding to a second geographical range into a first trained model obtained by machine learning using as input data information on the propagation path of radio waves from the transmission point to the reception point based on the arrangement of buildings in the first geographical range and as training data, to acquire an estimated value of the propagation path of radio waves from the transmission point to the reception point in the second geographical range; and a second acquisition means that inputs the estimated value of the propagation path obtained by inputting the input data corresponding to the first geographical range into the first trained model and inputs the estimated value of the propagation path of radio waves in the second geographical range acquired by the first acquisition means into a second trained model obtained by machine learning using as training data actual measured values ​​of the propagation characteristics of radio waves at the reception point in the first geographical range. [Effects of the Invention]

[0008] According to the present invention, it is possible to construct a model with high generalization performance and estimate radio wave propagation characteristics with high accuracy. [Brief explanation of the drawings]

[0009] [Figure 1] FIG. 1 illustrates an example of the configuration of an information processing device. [Figure 2] FIG. 1 is a diagram illustrating machine learning for estimating a propagation path. [Figure 3] FIG. 1 is a diagram illustrating machine learning for estimating a propagation path. [Figure 4] FIG. 10 is a diagram illustrating a procedure for estimating a propagation path using a trained model. [Figure 5] FIG. 10 is a diagram illustrating an example of a method for expressing a propagation path. [Figure 6] FIG. 1 is a diagram illustrating machine learning for estimating propagation characteristics. [Figure 7]FIG. 10 is a diagram illustrating a procedure for estimating propagation characteristics using a trained model. [Figure 8] FIG. 10 is a diagram illustrating an example of the processing flow of a learning phase. [Figure 9] FIG. 10 is a diagram illustrating an example of the processing flow of an estimation phase. DETAILED DESCRIPTION OF THE INVENTION

[0010] Hereinafter, the embodiments will be described in detail with reference to the accompanying drawings. Note that the following embodiments do not limit the scope of the invention as claimed, and not all combinations of features described in the embodiments are necessarily essential to the invention. Two or more of the features described in the embodiments may be combined in any desired manner. Furthermore, the same reference numerals are used to designate identical or similar components, and redundant descriptions will be omitted.

[0011] (Device configuration) In this embodiment, a propagation characteristic estimation system that estimates radio wave propagation characteristics using machine learning will be described. FIG. 1 is a diagram illustrating an example of the hardware configuration of an information processing device in which such a system is implemented. The information processing device is, for example, a general-purpose computer, and includes a processor 101, a storage device 102, an input / output interface 103, and an external storage device 104. The processor 101 includes one or more processors, such as a central processing unit (CPU) or a microprocessor unit (MPU). The processor 101 is configured to execute each process described below by executing a program stored in the storage device 102. The storage device 102 includes one or more memories, such as a read-only memory (ROM) or a random access memory (RAM). The storage device 102 stores, for example, programs corresponding to the following processes and variables used when the programs are executed. The input / output interface 103 is, for example, an interface for receiving input of information from outside the device and outputting information to outside the device. The input / output interface 103 includes an interface for accepting user operations using, for example, a keyboard or a pointing device, and presenting information to the user using a display or a speaker. The input / output interface 103 may also be a communication interface. For example, inputs such as image data and training data to be used in machine learning, and image data indicating the geographical situation of a target for estimating propagation characteristics, are accepted via a communication line connected by the input / output interface 103. For example, the results of estimating propagation characteristics may be provided to an external device via the communication line connected by the input / output interface 103. The external storage device 104 is a removable storage device such as a hard disk drive (HDD), a solid-state drive (SSD), or a universal serial bus (USB) memory, and can store information such as image data. The configuration of FIG. 1 may be realized, for example, by the processor 101 executing a program stored in the storage device 102.Then, using the configuration of Figure 1, machine learning is performed via the input / output interface 103 or using information stored in the external storage device 104 to generate a trained model for estimating propagation characteristics, and the trained model is used to estimate the propagation characteristics of radio waves.

[0012] (Propagation path estimation and propagation characteristic estimation using machine learning) In this embodiment, a trained model of the entire system is formed through two types of learning phases. In the first learning phase, image data is input to learn propagation paths such as radio wave reflection and diffraction. In this learning, the radio wave propagation paths between the radio wave transmission point and reception point estimated by ray tracing are used as training data. Thus, when image data corresponding to a specific geographical area is input, learning is performed to estimate how radio waves will propagate in that geographical area. Note that, since ray tracing is performed in this learning, a certain amount of calculation is required. However, once learning is completed, further ray tracing is no longer necessary, thereby significantly reducing the amount of calculation compared to performing ray tracing for the entire geographical area. In this first learning phase, physically meaningful features such as radio wave reflection and diffraction are extracted from the input image data. In this way, the first trained model obtained in the first learning phase outputs an information sequence indicating the propagation path in that geographical area in response to input of image data corresponding to the geographical area including the radio wave transmission point and reception point.

[0013] Then, a second learning phase is performed to output information on propagation characteristics such as propagation loss and spatiotemporal characteristics, using an information sequence indicating the propagation path obtained by inputting image data into the first trained model as input. In this second learning phase, actually measured propagation characteristics are used as training data. That is, machine learning is performed by using an information sequence indicating the propagation path obtained for reception points included in an image of a geographical range corresponding to the image data used in the first learning phase as input, and the propagation characteristics measured at those reception points as training data.

[0014] FIG. 2 is a diagram illustrating machine learning for estimating radio wave propagation paths. In this embodiment, machine learning is performed using elevation and building height data 201 related to the elevation and building heights in a predetermined geographical range and location information 202 of transmission and reception points in that geographical range. The elevation and building height data 201 includes, for example, information on two-dimensional map image data (e.g., satellite image data) showing the geographical range from above, and data indicating the elevation and building height information at positions corresponding to each pixel of the map image data. Note that each pixel of the two-dimensional map data may be represented by coordinates, for example, where the east-west direction is indicated by the x-coordinate and the north-south direction is indicated by the y-coordinate. Furthermore, one pixel corresponds to an area of ​​a predetermined size, such as an area having a width of 5 meters (m) in both the x- and y-directions. However, this is just an example, and the size of the geographical range corresponding to one pixel is determined according to the scale of the map data. Furthermore, the elevation and building height data 201 may include data on the building occupancy rate of each area within the geographical range corresponding to the map image data. Furthermore, the elevation / building height data 201 may include 2.5D or 3D shape data that indicates the shape of the building. The transmitting / receiving point location information 202 includes coordinates and height information of the point where radio waves are transmitted and the point where radio waves are received within the geographical range. In one example, the point where radio waves are transmitted is the location where the antenna of the base station device is installed, and the point where radio waves are received can be any location where a terminal device may be located.

[0015] In ray tracing 211, a three-dimensional model that reproduces the geographical situation in the geographical range is formed based on the elevation / building height data 201. For example, by placing objects with heights corresponding to the coordinates within the geographical range in a virtual space that includes the geographical range, a three-dimensional space model is formed that simulates the topography of the geographical range and objects that may reflect radio waves. Then, in ray tracing 211, when radio waves are transmitted from a transmitting point, how the radio waves will reach a receiving point after passing through reflection and diffraction is calculated. Note that ray tracing is a conventional technology, and will not be described in detail here. Such information calculated by ray tracing 211, which indicates the propagation path that the radio waves follow from the transmitting point to the receiving point, is called ray information 212.

[0016] The environmental information generation 221 may generate, for example, building height data 222 based on the height of the transmission point, building height data 223 based on the height of the reception point, and building occupancy data 224. The building height data 222 based on the height of the transmission point is, for example, information indicating the height of buildings in an area corresponding to each pixel of two-dimensional map image data when the geographical range is viewed from above, based on the height of the transmission point. For example, if the height z1(m) of the transmission point is a pixel value (e.g., brightness value) = 128, the height of buildings in an area corresponding to each pixel is expressed by the pixel value based on the relative height from that height. If the height of a building in an area corresponding to the i-th pixel of the map image data is z2(m), the pixel value p(i) of that pixel may be expressed as p(i) = 128 + (z2 - z1). That is, if the height z1 of the transmission point is 30 m and the height of the building in the area corresponding to the i-th pixel is 50 m, the pixel value can be determined as p(i) = 128 + (50 - 30) = 148. Similarly, for building height data 223 based on the height of the reception point, if the height of the reception point is z3 and the height of the building in the area corresponding to the i-th pixel is z2 (m), the pixel value p(i) of that pixel can be expressed as p(i) = 128 + (z2 - z3). In this way, building height data 222 based on the height of the transmission point and building height data 223 based on the height of the reception point can be output as image data of the same size as the map image data. Note that this is just an example, and the building heights in each area based on the height of the transmission point or the reception point may be specified in other formats. Building occupancy data 224 is, for example, information indicating the proportion of the area covered by buildings in the land area of ​​an area corresponding to one pixel of map image data. For example, the degree to which the area corresponding to each pixel is covered by buildings can be indicated by setting the pixel value for a building occupancy of 0% to 0, the pixel value for 50% to 128, and the pixel value for 100% to 255. In this way, the building occupancy data 224 can also be output as image data of the same size as the map image data. Note that the format of this data is not limited to this, and the building occupancy data 224 may also be prepared in another format.The data generated as environmental information is not limited to the data 222 of building heights based on the height of the transmission point, the data 223 of building heights based on the height of the reception point, and the data 224 of building occupancy rates, and other data may be generated. Also, only one or two of the data 222 of building heights based on the height of the transmission point, the data 223 of building heights based on the height of the reception point, and the data 224 of building occupancy rates may be generated.

[0017] The information generated as described above is then used for machine learning 231. In this embodiment, the ray information 212 generated by the ray tracing 211 is used as training data 213, and the information generated by the environment information generation 221 is used as input data 225 to perform machine learning 231. Here, machine learning may be performed by, for example, a neural network using a Long Short-Term Memory (LSTM). However, this is just one example, and other machine learning techniques such as a Gated Recurrent Unit (GRU), a graph neural network, or reinforcement learning may also be used. In the first learning phase, various data generated in the environment information generation 221 is input to a learning model to obtain an output. Then, the output is compared with training data 213 (ray information 212 calculated for the corresponding environment), and learning is performed so that the difference between them becomes sufficiently small. That is, the learning model is updated using various combinations of the elevation / building height data 201 and the position information 202 of the transmitting and receiving points, while feeding back the difference between the training data 213 and the output of the learning model. The first learning phase is completed when the difference between the inputs converges to a sufficiently small value, and the learning model at the time of completion of the learning can then be used as a trained model to estimate the propagation path. The configuration of the machine learning is a conventional technique, and therefore further details will not be described here.

[0018] As described above, in the first learning phase of this embodiment, ray information generated using ray tracing is used as training data to perform machine learning for propagation path estimation using information on building heights and occupancy rates corresponding to map image data as input. While the above example describes an example in which machine learning is performed using data generated in the environmental information generation 221 as input, this is not limiting. For example, as shown in FIG. 3 , feature extraction 301 may be performed on the data generated in the environmental information generation 221 before the data is used as input data. This feature extraction 301 may be performed in a manner similar to the feature extraction of an image by convolution in a convolutional neural network. Performing feature extraction in this manner enables more efficient generation of a trained model for propagation path estimation using characteristic elements of the data generated in the environmental information generation 221.

[0019] FIG. 4 is a diagram illustrating propagation path estimation after a trained model is obtained. In propagation path estimation, for example, environment information generation 221 is performed using a combination of elevation and building height data 201 and transmitting / receiving point location information 202 corresponding to a geographical range different from the information used in the learning phase in FIG. 2. This environment information generation 221 is performed using the same procedure as in the learning phase. Furthermore, as shown in FIG. 3, if feature extraction is performed in the learning phase, similar feature extraction is also performed in propagation path estimation using the trained model to generate input data. As a result, data in the same format as the input data in the learning phase is generated. This data is then input as input data to a trained model 401. The output of the trained model 401 becomes a propagation path estimate value 402.

[0020] As described above, in this embodiment, in the first learning phase, the trained model is configured to output propagation path estimates for image-format data such as building height data 222 based on the height of a transmission point corresponding to map image data, building height data 223 based on the height of a reception point, and building occupancy data 224. The propagation path estimates may take the form of a sequence of coordinates of reflection and diffraction points, for example. For example, if the coordinates of the transmission point of a radio wave are (x0, y0), the radio wave is reflected at a position with coordinates (x1, y1) and a position with coordinates (x2, y2), and reaches the radio wave reception point represented by coordinates (x3, y3), a sequence such as {(x0, y0), (x1, y1), (x2, y2), (x3, y3)} may be output as the propagation path estimates. Alternatively, IDs may be assigned to reflection surfaces and diffraction edges, and a sequence using the IDs may be output as the propagation path estimates. For example, a sequence of information such as {reflection at the surface of ID1, diffraction at the edge of ID2, ...} may be output as a propagation path estimate. Furthermore, regardless of the presence or absence of reflection or diffraction, the propagation path may be divided into unit distances, and the propagation path estimate may be determined by a sequence of coordinates of the division positions. For example, as shown in FIG. 5, the propagation path from the radio wave transmission point (x1, y1) to the radio wave reception point (xn, yn) may be divided into unit distances, and the propagation path may be expressed using the coordinates of the division positions (x2, y2), (x3, y3), ..., as {(x1, y1), (x2, y2), (x3, y3), ... (xn, yn)}. Furthermore, a sequence of propagation direction vectors [v1], [v2], [v3], ... at the division positions may be used to represent the propagation path. Note that the sequence of such information may be a variable-length sequence. When a variable-length sequence is to be output, machine learning is performed using a technique capable of handling a learning model corresponding to such variable-length output. In this embodiment, for example, a recurrent neural network (RNN) such as LSTM or GRU, a graph neural network (GNN), reinforcement learning, or the like is used to enable handling of variable-length propagation path information. However, this is merely an example, and the propagation path information may be expressed as fixed-length information. Furthermore, the propagation path information may be output in a format different from the above format.

[0021] In the example of the first learning phase described above, ray information obtained by ray tracing 211 is used as training data, but this is not limiting. For example, spatiotemporal characteristics may be measured at a specific receiving point, and the measured values ​​may be used as training data. Examples of spatiotemporal characteristics include an angle profile and a delay profile. Spatiotemporal characteristics may be measured by transmitting radio waves multiple times from a transmitting point and measuring the radio waves while changing the directivity of a receiving antenna at the receiving point. At this time, clustering may be performed, in which a group of measurement data that is assumed to have arrived via the same path in the multiple measurement results is treated as one cluster. Then, spatiotemporal characteristics indicating angle information or delay information of the cluster may be used as training data.

[0022] Once the propagation path estimates are obtained in the first learning phase, a second learning phase is executed to estimate propagation characteristics such as propagation loss and spatiotemporal characteristics based on the propagation path estimates. In the second learning phase, as shown in FIG. 6 , for example, machine learning 631 is performed using propagation characteristic measurement values ​​601 as training data 602 and propagation path estimate values ​​611 estimated using the trained model obtained in the first learning phase as input data 612. If the propagation path estimates are variable-length data, for example, an RNN 613 may be used to extract features from the variable-length data and output fixed-length data. The RNN 613 is just an example, and other techniques capable of extracting features from variable-length data and converting the data into fixed-length data may be used. If the propagation path estimates are fixed-length data or if variable-length data can be handled in the machine learning 631, the RNN 613 may be omitted. The propagation characteristics are estimated using, for example, a fully connected neural network (FNN). In machine learning 631, input data 612 is input to a neural network, and the resulting propagation characteristic estimates are compared with training data 602, and learning is performed so that the difference between them becomes sufficiently small. The propagation characteristic values ​​measured at the reception point of the map image data used to obtain the propagation path estimates 611 used as the input data 612 are used as training data 602. The learning model is updated by feeding back the difference between the training data 602 and the output obtained by inputting the propagation path estimates 611 into the learning model. The second learning phase is completed when the difference between the inputs of various propagation path estimates 611 has converged to a sufficiently small state. After that, the learning model at the time of completion of learning can be used as a trained model to estimate propagation characteristics. The configuration of machine learning is a conventional technique, and therefore further details will not be described here.

[0023] FIG. 7 is a diagram illustrating the estimation of propagation characteristics after a trained model is obtained. In the estimation of propagation characteristics, for example, a propagation path estimate 611 obtained using the configuration shown in FIG. 4 corresponding to a geographical area different from the information used in the learning phase in FIG. 6 is input to the trained model 701 as input data 612. Note that if the variable-length data of the propagation path estimate 611 is converted to fixed-length data using an RNN 613 or the like in the second learning phase, a process similar to the conversion process may be performed on the propagation path estimate 611 before inputting it to the trained model 701. In other words, data of the same format is used during learning and estimation. The output of the trained model 701 becomes a propagation characteristic estimate 702.

[0024] (Processing flow) An example of the flow of machine learning processing in this embodiment is shown in Fig. 8. Note that the specific content of each process has been described above, so only an outline of the processing flow will be given here.

[0025] In this process, ray tracing is first performed based on elevation and building height data and location information of the transmitting and receiving points (S801), generating training data for machine learning related to propagation path estimation. In parallel with this, input data for machine learning, which is also used as input data for propagation path estimation, is generated based on the elevation and building height data and location information of the transmitting and receiving points (S802). Then, machine learning is performed using the training data generated in S801 and the input data generated in S802 (S803). This procedure is repeatedly performed using information about multiple geographical areas, thereby acquiring a first trained model for propagation path estimation. At this point, data in a format similar to the input data generated in S802 is input into the first trained model, thereby estimating the radio wave propagation path from the transmitting point to the receiving point specified when the input data was generated. The processing up to this point corresponds to the first learning phase described above.

[0026] Thereafter, for example, the input data generated in S802 is input to the first trained model, and an estimated value of the propagation path is acquired (S804). The acquired estimated value of the propagation path is treated as input data in machine learning for estimating the propagation characteristics. In addition, actual measured values ​​of the propagation characteristics, which are the results of measuring radio waves transmitted from a transmitting point at a receiving point in a geographical range corresponding to the elevation and building height data used when the input data was generated in S802, are used as training data. In this way, machine learning is performed using the estimated value of the propagation path as input and the values ​​of the actually measured propagation characteristics in the environment in which the propagation path is obtained as training data (S805). By repeatedly executing this procedure using multiple estimated values ​​of the propagation paths and corresponding training data, a second trained model for estimating the propagation characteristics is acquired. The processes of S804 and S805 correspond to the second learning phase described above.

[0027] FIG. 9 shows an example of the flow of a process for estimating propagation characteristics based on elevation / building height data and location information of transmitting and receiving points, which is performed after the first trained model and the second trained model are acquired. In this process, input data is generated based on elevation / building height data and location information of transmitting and receiving points (S901). This process corresponds to the process of S802. However, since machine learning has been completed here, no training data is required. Therefore, ray tracing is not performed. Then, for example, from map image data of a geographical area where ray tracing and propagation characteristics measurements have not been performed, data of the type used for machine learning is generated from among building height data based on the transmitting point, building height data based on the receiving point, and building occupancy data. Then, the data generated in S901 is input to the first trained model (S902). As a result, an estimated value of the propagation path is obtained. Thereafter, the estimated value of the propagation path is input to the second trained model (S903). As a result, an estimated value of the propagation characteristic is obtained.

[0028] As described above, in this embodiment, a first trained model and a second trained model are acquired, and then the first trained model and the second trained model are linked to generate a single model. As a result, a propagation path is estimated from image data that indicates, as a single pixel, the building height or building occupancy rate in each of multiple areas of a predetermined size included in a predetermined geographical area, and propagation characteristics are estimated from the estimated propagation path. As described above, in this embodiment, physically meaningful features related to radio wave propagation are extracted from the image data, and propagation characteristics are estimated using the features. This makes it possible to build a model with high generalization performance and estimate radio wave propagation characteristics with high accuracy. Therefore, it is possible to contribute to Goal 9 of the United Nations' Sustainable Development Goals (SDGs), "Build resilient infrastructure, promote sustainable industrialization and foster innovation."

[0029] The invention is not limited to the above-described embodiment, and various modifications and variations are possible within the scope of the gist of the invention.

Claims

1. a first acquisition means for acquiring a first trained model by machine learning using, as input data, image data indicating, as one pixel, the height or occupancy rate of buildings in each of a plurality of regions of a predetermined size included in a predetermined geographical range, and using, as training data, information on the propagation path of radio waves from a transmission point to a reception point based on the layout of buildings in the predetermined geographical range; a second acquisition means for acquiring a second trained model by machine learning using an estimated value of a propagation path obtained by inputting the image data into the first trained model as an input and an actual measurement value of the propagation characteristics of the radio wave at the receiving point as training data; a generation means for linking the first trained model and the second trained model to generate a model for estimating radio wave propagation characteristics from the image data; An information processing device comprising:

2. 2. The information processing device according to claim 1, wherein the first acquisition means acquires training data by ray tracing based on information on the height and location of buildings in the specified geographical range and position information of the transmission points and the reception points.

3. 2. The information processing device according to claim 1, wherein the first acquisition means performs the machine learning using a neural network using a Long Short-Term Memory (LSTM) or a Gated Recurrent Unit (GRU), a graph neural network, or reinforcement learning.

4. 4. The information processing device according to claim 3, wherein the estimated value of the propagation path obtained by inputting the image data into the first trained model is represented by a variable-length sequence.

5. 5. The information processing apparatus according to claim 4, wherein said second obtaining means further includes converting means for converting a variable-length sequence input as said propagation path estimate into a fixed-length sequence.

6. a first acquisition means for inputting second image data in the same format as the input data corresponding to a second geographical range into a first trained model acquired by machine learning using as input data first image data indicating, as one pixel, the height or occupancy rate of buildings in each of a plurality of regions of a predetermined size included in a first geographical range, and using as training data information on the propagation path of radio waves from transmission points to reception points based on the layout of buildings in the first geographical range; and a second acquisition means for inputting an estimated value of a propagation path obtained by inputting the input data corresponding to the first geographical range into the first trained model as an input, and inputting the estimated value of the propagation path of the radio wave in the second geographical range acquired by the first acquisition means into a second trained model acquired by machine learning using actual measured values ​​of the propagation characteristics of the radio wave at the reception points in the first geographical range as training data, thereby acquiring the propagation characteristics at the reception points in the second geographical range; An information processing device comprising:

7. the estimated value of the propagation path obtained by inputting the second image data into the first trained model is represented by a variable-length sequence; 7. The information processing device according to claim 6, wherein, when the second trained model is acquired by machine learning using, as input data, data in which a variable-length sequence input as an estimate of a propagation path of the radio wave in the first geographical range is converted into a fixed-length sequence, the second acquisition means further includes conversion means for converting a variable-length sequence representing an estimate of a propagation path of the radio wave in the second geographical range into a fixed-length sequence.

8. An information processing method executed by an information processing device, acquiring a first trained model by machine learning using image data indicating, as input data, the building height or building occupancy rate in each of a plurality of regions of a predetermined size included in a predetermined geographical range as one pixel, and using, as training data, information on the propagation path of radio waves from a transmission point to a reception point based on the arrangement of buildings in the predetermined geographical range; acquiring a second trained model by machine learning using an estimated value of a propagation path obtained by inputting the image data into the first trained model as an input and an actual measurement value of the propagation characteristics of the radio wave at the receiving point as training data; Linking the first trained model and the second trained model to generate a model that estimates radio wave propagation characteristics from the image data; and An information processing method comprising:

9. An information processing method executed by an information processing device, inputting second image data in the same format as the input data corresponding to a second geographical range into a first trained model obtained by machine learning using as input data first image data indicating, as one pixel, the height or occupancy rate of buildings in each of a plurality of regions of a predetermined size included in a first geographical range, and using as training data information on the propagation path of radio waves from transmission points to reception points based on the layout of buildings in the first geographical range; and acquiring an estimate of the propagation path of radio waves from transmission points to reception points in the second geographical range; inputting the estimated value of the propagation path obtained by inputting the input data corresponding to the first geographical range into the first trained model as an input, and inputting the estimated value of the propagation path of the radio waves in the second geographical range into a second trained model acquired by machine learning using actual measured values ​​of the propagation characteristics of the radio waves at the reception points in the first geographical range as training data, thereby acquiring the propagation characteristics of the radio waves at the reception points in the second geographical range; An information processing method comprising:

10. A program for causing a computer to function as each of the means included in the information processing device according to any one of claims 1 to 7.

Images