JP7369402B2 - Learning methods, learning devices and programs - Google Patents

Description

The present invention relates to a learning method, a learning device, and a program.

Expectations are high for millimeter wave communication as a next-generation wireless communication technology that can realize large-capacity and high-speed communication (see, for example, Non-Patent Document 1). One of the advantages of millimeter wave communication is that the available frequency width is wide, and high-speed communication exceeding 1 Gbit/s (gigabits per second) is possible. On the other hand, since millimeter waves are highly attenuated by moisture and oxygen, they have the disadvantage that communication quality drops rapidly if the line-of-sight communication path is blocked by a human body or the like (for example, see Non-Patent Document 2).

In order to solve the line-of-sight channel shielding problem in millimeter-wave communication, a handover system for millimeter-wave communication using an RGB-D (RGB and Depth) camera is being considered (for example, see Non-Patent Documents 3 and 4). . This system uses an RGB-D camera to observe the millimeter wave communication environment, and estimates and predicts the current and future millimeter wave communication environment from the image and video data captured by the camera. By performing handover control based on estimated and predicted data, it is possible to avoid deterioration in communication quality due to changes in the millimeter wave communication environment, including line-of-sight channel blocking, without compressing the wireless band with control signals. Furthermore, by estimating communication quality information between a terminal and a base station that is not in communication from a depth image, it is possible to obtain an index for selecting a base station to switch to without consuming wireless bands. Non-Patent Document 4 proposes millimeter wave communication quality estimation using an RGB-D camera and machine learning. In the technique of Non-Patent Document 4, a deep neural network is used to learn a correspondence model between depth images obtained from an RGB-D camera and received signal power, and the learned model is used to calculate received signal power from a new depth image. Estimate.

Furthermore, the possibility of learning the correspondence between depth images and MCS (Modulation and Coding Scheme) indexes has been studied (for example, see Non-Patent Document 5). MCS is generally expressed as an MCS index table that stores combinations of modulation schemes and coding rates. By determining the MCS index, the corresponding modulation method and coding rate are determined. Determining the MCS index is also referred to as MCS control. Non-Patent Document 5 shows that, assuming that the optimal MCS for a certain image is known, a model for estimating an MCS index from a depth image can be learned using a deep neural network.

Furthermore, there is a conventional technique that performs MCS control without using a camera image (for example, see Non-Patent Document 6). In this technique, the next MCS index is determined only from past communication history. As a method for determining the MCS index, there is a method in which the MCS index is increased when communication is successful a certain number of times, and is decreased when communication fails. However, these implementations are vendor dependent and are not standardized.

Peng Wang, Yonghui Li, Lingyang Song, Branka Vucetic, "Multi-Gigabit Millimeter Wave Wireless Communications for 5G: From Fixed Access to Cellular Networks", IEEE Communications Magazine, January 2015, p. 168-178 Sylvain Collonge, Gheorghe Zaharia, Ghais El Zein, "Influence of the Human Activity on Wide-Band Characteristics of the 60 GHz Indoor Radio Channel", IEEE Transactions on Wireless Communications, November 2004, vol.3, no.6, p. .2396－2406 Yuta Oguma, Ryohei Arai, Takayuki Nishio, Koji Yamamoto, Masahiro Morikura, "Proactive Base Station Selection Based on Human Blockage Prediction Using RGB-D Cameras for mmWave Communications", 2015 IEEE Global Communications Conference (GLOBECOM), 2015, p.1 -6 Takayuki Nishio, Hironao Okamoto, Kota Nakashima, Yusuke Koda, Koji Yamamoto, Masahiro Morikura, Yusuke Asai, Ryo Miyatake, "Proactive Received Power Prediction Using Machine Learning and Depth Images for mmWave Networks", IEEE Journal on Selected Areas in Communications, Vol. 37, No.11, November 2019 Tomoya Mikuma, Satoshi Nishio, Masahiro Morikura, Yusuke Asai, Ryo Miyatake, "[Poster Presentation] Examination of MCS Index prediction based on depth images for transmission rate control of millimeter wave communications", Electronic Information and Communications Corporation Academic Society, IEICE Technical Report, January 2020, SeMI2019-111, p.51-52 Dong Xia, Jonathan Hart, Qiang Fu, "Evaluation of the Minstrel Rate Adaptation Algorithm in IEEE 802.11g WLANs", IEEE ICC 2013 - Communication QoS, Reliability and Modeling Symposium, 2013, p.2223-2228

In Non-Patent Document 5, the possibility of mapping from an image to MCS is studied. FIG. 8 is a diagram showing the process flow of Non-Patent Document 5. This technique assumes that in a situation where a base station and a terminal are communicating using millimeter waves, a data set is prepared in advance to which an optimal MCS index that is considered to be the correct answer for a depth image is assigned. Then, using the data set, a deep neural network learns mapping from the depth image to the MCS index. Non-Patent Document 5 shows using experimental data that by using a learned model, it is possible to directly estimate the optimal (supposedly) MCS index when a new depth image is obtained. This shows that predictive MCS control can reduce the line disconnection probability.

Conventional MCS control methods that do not use camera images, as in Non-Patent Document 6, cannot cope with a steep drop in received signal power in millimeter wave communication, and change the MCS after a large amount of packet loss occurs. Therefore, communication quality deteriorates in an environment where the communication path is frequently blocked. In the above-mentioned Non-Patent Document 5, MCS control is performed using camera images. However, in Non-Patent Document 5, the main focus is on verifying whether the correspondence between depth images and MCS index can be learned, and the data set of depth images and MCS used for learning is not considered.

In view of the above circumstances, the present invention provides a learning method that can learn a modulation method and coding rate estimation model that reduces burst errors in wireless communication by using prepared learning data while reducing manual effort. The purpose is to provide methods, learning devices and programs.

One aspect of the present invention provides an image captured of a communication environment of a wireless communication system, and a modulation method and coding rate, or a modulation method and coding rate that were used when wireless communication was successful in the wireless communication system. This learning method includes a learning step of learning a model representing a correspondence relationship between an image and an index value based on data in which the image and the index value are associated with each other.

One aspect of the present invention provides an image captured of a communication environment of a wireless communication system, and a modulation method and coding rate, or a modulation method and coding rate that were used when wireless communication was successful in the wireless communication system. This learning device includes a learning unit that learns a model representing a correspondence relationship between an image and an index value based on data in which the image and the index value are associated with each other.

One aspect of the present invention is a program for causing a computer to function as the above-mentioned learning device.

According to the present invention, it is possible to learn a modulation scheme and coding rate estimation model that reduce burst errors in wireless communication using prepared learning data while reducing manual effort.

1 is a configuration diagram of a wireless communication system according to a first embodiment of the present invention. It is a figure which shows the example of the MCS index table used for the same embodiment. FIG. 2 is a flow diagram showing processing of the wireless communication system according to the embodiment. FIG. 7 is a diagram illustrating an example of penalty asymmetry at the time of MCS misestimation according to the same embodiment. FIG. 3 is a diagram showing line disconnection probabilities in the prior art and the embodiment. It is a figure which shows the average throughput of each of a prior art and an embodiment. FIG. 2 is a diagram showing a hardware configuration according to an embodiment. FIG. 2 is a diagram showing a process flow of a conventional technique.

Embodiments of the present invention will be described in detail below with reference to the drawings.

(First embodiment)
FIG. 1 is a configuration diagram of a wireless communication system according to a first embodiment of the present invention. The wireless communication system 1 includes a base station (BS) 2, a camera 3, a learning device 4, and a wireless terminal (STA) 5.

The base station 2 performs millimeter wave communication with the STA 5 that the user uses while moving. The base station 2 and STA 5 are examples of wireless communication devices. While millimeter wave communication can achieve high-capacity and high-speed communication, the line-of-sight communication path is blocked by human bodies, etc., resulting in a sharp decline in communication quality. In order to solve this problem, the learning device 4 uses the learned model to predictively determine an MCS (Modulation and Coding Scheme) index from the camera image.

As described above, in millimeter wave communication, the desired signal is greatly attenuated when shielding occurs, resulting in a decrease in signal to noise ratio (SNR). Therefore, the learning device 4 avoids packet burst loss by predicting the occurrence of shielding in advance and selecting an MCS index that allows successful communication even when the SNR is low. Specifically, the learning device 4 learns an MCS estimation model that directly estimates the optimal MCS index from images captured by the camera 3 that monitors the communication environment. For example, the technique of a reception power prediction method using camera images (Non-Patent Document 4) can be applied to the learning. It is assumed that the larger the value of the MCS index, the more the combination of modulation method and coding rate corresponds to a higher SNR and faster communication speed.

For learning, the learning device 4 stores the results of communication between the BS 2 and the STA 5 through conventional MCS control that does not use camera images in a communication history database together with the camera images at that time. Next, the learning device 4 calculates the similarity of each camera image stored in the communication history database, and calculates the similarity from the MCS index corresponding to the camera image whose similarity is higher than a predetermined value and whether or not communication was successful at that time. Find the maximum MCS index at which communication is successful. The learning device 4 labels all of those camera images whose similarity is higher than a predetermined value with the obtained maximum MCS index. The learning device 4 performs supervised learning of the correspondence between this camera image and the labeled MCS index using a neural network or the like. Thereby, the learning device 4 can estimate the MCS index with the highest transmission rate among the MCS indexes at which communication is successful even when an unknown camera image is obtained. According to this embodiment, MCS control that predicts future communication quality is possible, so burst errors can be reduced and high throughput can be achieved compared to conventional MCS control.

The configuration of each device will be explained.
The BS2 includes an RF (Radio Frequency) processing unit 21 that performs wireless communication with the STA5. The RF processing section 21 includes a communication control section 22 and an MCS control section 23. The communication control unit 22 performs millimeter wave communication with the STA 5 similarly to the conventional technology. The MCS control unit 23 instructs the communication control unit 22 to perform communication based on the MCS index notified from the learning device 4. The communication control unit 22 performs millimeter wave communication using a combination of modulation method and coding rate that corresponds to the MCS index instructed by the MCS control unit 23. Note that the MCS control unit 23 determines the MCS index in the same manner as in the prior art until the learning device 4 finishes learning the MCS estimation model.

The camera 3 is an example of an imaging device that captures a depth image. For example, the camera 3 is an RGB-D (RGB and Depth) camera, and captures camera images including distance information such as RGB-D images. The camera 3 takes a camera image including the wireless line-of-sight channel between the BS 2 and the STA 5 and its surroundings, and transmits it to the learning device 4 . The camera image may be a still image or a moving image. In the case of a moving image, each frame constituting the moving image is one camera image.
The camera image includes at least depth (distance) data for each pixel.

The learning device 4 includes an information management section 41, a learning data generation section 42, a machine learning section 43, and an estimation section 44. The information management unit 41 stores a communication history database. The communication history database holds a camera image captured by the camera 3 in association with an MCS index when the BS 2 communicated with the STA 5 at the time when the camera image was captured. In the communication history database, a camera image and an MCS index are set in association with each other when communication between the BS2 and the STA5 is successful. Note that the camera image, the MCS index, and information regarding whether or not the communication between the BS2 and the STA5 was successful may be set in the communication history database. The learning data generation unit 42 generates learning data used for model learning using a set of a camera image stored in the communication history database and an MCS index at the time of successful communication. The machine learning unit 43 uses the learning data generated by the learning data generating unit 42 to learn an MCS estimation model representing the correspondence between camera images and MCS index values. The estimation unit 44 uses the MCS estimation model learned by the machine learning unit 43 to determine the value of the MCS index to be used next by the BS 2 from the camera image, and instructs the determined MCS index to the BS 2.

Note that the learning device 4 may use all of the images generated from still images and videos taken by the camera 3 for machine learning of the MCS estimation model, and may arbitrarily select an image from among the generated images to perform the MCS It may also be used for machine learning of estimation models. In addition, the learning device 4 extracts a plurality of images from the generated images according to norms such as time series, and performs image processing such as averaging and difference extraction for each pixel, and converts the newly generated image into MCS. It may also be used for machine learning of estimation models.

FIG. 2 is a diagram showing an example of an MCS index table used in this embodiment. FIG. 2 shows an MCS index table in IEEE802.11 ad. The MCS index table shows the modulation method (Modulation) and coding rate (Code Rate) corresponding to each value of the MCS index. The value of MCS index is also written as MCS index number. Furthermore, the MCS index table includes N _CBPS (number of code bits per symbol), number of repetitions, and data rate corresponding to the MCS index number. As shown in FIG. 2, the larger the MCS index number, the larger the data rate.

FIG. 3 is a flow diagram showing the processing of the wireless communication system 1. The communication control unit 22 of the BS2 communicates wirelessly with the STA5 (step S105). The communication control unit 22 outputs the time information and the MCS index used for communication with the STA 5 at that time to the learning device 4 in association with each other (step S110). The communication control unit 22 may output the MCS index to the learning device 4 only when the communication is successful, or may add information on whether the communication is successful or not to the MCS index and output it to the learning device 4. good. To determine whether or not communication is successful, arbitrary indicators and thresholds can be used, such as the fact that the occurrence of errors is less than or equal to a predetermined value. On the other hand, the camera 3 outputs the captured camera image to the learning device 4 in real time (step S115). Camera images are associated with time information.

The information management unit 41 of the learning device 4 associates the camera image obtained from the camera 3 with the MCS index when the BS 2 communicated at the time indicated by the camera image in the communication history database, and stores the linked image in the communication history database. (Step S120). Note that while the MCS estimation model is not fully trained, the BS 2 communicates with the STA 5 using conventional rate control (for example, see Non-Patent Document 6) without using camera images, and the information management unit 41 maintains the MCS index of its rate control results. Here, it is assumed that the information management unit 41 holds only data that resulted in successful communication.

The learning data generation unit 42 uses the communication history database stored in the information management unit 41 to create teacher data for learning the MCS estimation model (step S125). The training data is learning data in which correct answers are given. The specific training data generation procedure will be described later. The machine learning unit 43 learns the MCS estimation model using the teacher data generated by the learning data generation unit 42 (step S130). As the MCS estimation model, a deep neural network including convLSTM (Convolutional Long Short-Term Memory), which is used in Non-Patent Documents 4 and 5, can be considered. This deep neural network can perform highly accurate learning when camera images are input, and can learn the relationship between output and time-series input data.

For example, in Non-Patent Document 4, depth images obtained from each of a plurality of past camera images are compressed into an image of a predetermined size. The deep neural network of Non-Patent Document 4 receives the plurality of compressed depth images as input, and outputs the power of a predetermined time period ahead of the newest depth image among the plurality of depth images. Instead of this power, the MCS estimation model of this embodiment outputs an MCS index that is a predetermined time α (α≧0) after the newest depth image among the plurality of depth images used for input.

After the MCS estimation model is learned, the estimation unit 44 estimates an MCS index suitable for the communication environment at that time by inputting the newly obtained camera image from the camera 3 into the learned MCS estimation model. (Step S135). The estimation unit 44 notifies the determined MCS index to the BS2.

The MCS control unit 23 of the BS 2 determines to use the MCS index notified from the learning device 4 in the next communication period (step S140). The MCS control unit 23 instructs the communication control unit 22 to perform communication using the determined MCS index. The communication control unit 22 performs millimeter wave communication with the STA 5 using the modulation method and coding rate corresponding to the instructed MCS index. When the communication control unit 22 ends communication with the STA 5 (step S145), the wireless communication system 1 repeats the processing from step S110. Note that the wireless communication system 1 may repeat the processes from step S110 to step S140 while the BS 2 is communicating with the same STA 5.

Next, a specific procedure for generating teacher data in the learning data generating section 42 will be described. The learning data generation unit 42 classifies combinations of camera images and MCS indexes read from the information management unit 41 based on the similarity of the camera images. Specifically, the learning data generation unit 42 calculates the distance between the two camera images for all combinations of the two camera images, and uses the distance as the degree of similarity. That is, the learning data generation unit 42 classifies camera images whose distance is closer than a predetermined value as similar images. An example of distance is the Euclidean norm of depth between camera images. Alternatively, the learning data generation unit 42 may input the camera image to the received power prediction model of Non-Patent Document 4, and use the difference between the estimated received power values as the distance.

Next, the learning data generation unit 42 searches for the largest MCS index number among the data groups of camera images classified as similar images, and applies the searched MCS index number to all camera images classified into that data group. Label the MCS index number as the correct answer. The machine learning unit 43 uses the thus labeled camera images as teacher data for model learning.

Next, a loss function used when the machine learning unit 43 performs machine learning on the MCS estimation model will be described. In machine learning, parameter values used in the model are changed so that the output obtained by inputting teaching data into the model being trained approaches the correct answer indicated by the teaching data. A value calculated by a loss function is used as an index of how close the model output is to the correct answer. In the present embodiment, the machine learning unit 43 uses a loss function to calculate the error between the estimation result obtained by inputting the camera image of teacher data into the MCS estimation model and the correct answer indicated by the teacher data. The machine learning unit 43 changes the parameter values used in the MCS estimation model such as a deep neural network so that the error calculated by the loss function becomes smaller. Regarding the loss function, for example, see References: Yasuki Saito, "Deep Learning from Scratch - Learning Theory and Implementation of Deep Learning with Python", 4.2 Loss Function, Oilery Japan Co., Ltd., 2016, p.87- 91”.

In machine learning of the MCS estimation model, a method of setting a loss function may be considered, focusing on the asymmetry of the penalty when the MCS index is incorrectly estimated. The asymmetry of the penalty when the MCS index is incorrectly estimated is as follows. That is, in an environment where communication is actually possible only with a low value MCS index, if a high value MCS index is erroneously selected, packet burst errors and the like occur, resulting in a significant deterioration of communication quality. Conversely, if a low value MCS index is erroneously selected in an environment where communication is actually possible with a high value MCS index, no significant deterioration in communication quality will occur, and only a decrease in communication speed will occur. In this way, the asymmetry of the penalty when the MCS index is incorrectly estimated represents the asymmetry of the amount of quality deterioration.

FIG. 4 is a diagram illustrating an example of penalty asymmetry at the time of MCS misestimation. Assume that the true value of the maximum MCS index that can be used in the current environment is "5". If the value of the MCS index estimated by the MCS estimation model is "5", the estimation is correct and communication is performed at the maximum speed currently possible. On the other hand, if the value of the MCS index estimated by the MCS estimation model is "7", this is an erroneous estimation and a communication burst error occurs. Further, if the value of the MCS index estimated by the MCS estimation model is "2", the communication quality will not deteriorate, but communication will be performed at a speed lower than the currently possible maximum speed.

In wireless communications, a decrease in communication speed has a smaller impact on communication quality than communication failure due to burst errors. Therefore, even if erroneous estimation occurs, it is desirable to erroneously estimate the MCS index to be a value lower than the true MCS index. Therefore, in machine learning, we focus on the magnitude relationship between the true value and estimated value of the MCS index, and when the estimated value is larger than the true value, we increase the output of the loss function, and conversely, we increase the output of the loss function when compared with the true value. If the estimated value is small, the output of the loss function is reduced. This performs learning to prevent burst errors.

The following describes how to set the loss function in the conventional case in which the asymmetry of penalties is not reflected and in the case in which the asymmetry is reflected. First, a conventional case that does not reflect penalty asymmetry will be explained. As an example of a loss function, a sum-of-squares error formula widely used in machine learning will be used for explanation. Equation (1) shows the equation for the sum of squares error E.

Here, K is the number of elements in the y array and t array, and k is the element number representing the index of the array. That is, y _k represents the kth element of the y array, and t _k represents the kth element of the t array. The y array is an array that stores the probability that the input signal to machine learning is the MCS index of each value. Also, let the element number with the largest value in the t array be i, and the element number with the largest value in the y array be j. The t array is a one-hot representation of training data (=true value).

A specific example of the y array, t array, element number i, and element number j will be shown. For example, consider a problem in which the candidate value of MCS index is 1, 2, 3, 4, or 5, and the true value is MCS index=2. First, since there are five candidate values for the MCS index, the number of elements K=5. At this time, the t array is a one-hot expression in which 1 is set only to the element with the element number corresponding to the true value, and 0 is set to the elements with other element numbers, so t = [0 1 0 0 0]. Therefore, for example, the value of t _k when k=2 is t ₂ =1, and the value of t _k when k=3 is t ₃ =0.

Next, the y array is an array of estimated values that stores the probability that the input signal has each value of the MCS index. That is, the y array is obtained from the output when inputting the input signal to the MCS estimation model. The probability that MCS index = 1 is 0.1, the probability that MCS index = 2 is 0.8, the probability that MCS index = 3 is 0.02, the probability that MCS index = 4 is 0.03, MCS index If the probability that =5 is 0.05, then y=[0.1 0.8 0.02 0.03 0.05]. Further, the maximum value in the t array is the value 1 of the second element, so i=2, and the maximum value in the y array is the value 0.8 of the second element, so j=2.

In equation (1), whether the value of element number j obtained from the estimated MCS index is lower or higher than the value of element number i obtained from the true MCS index, the element number When the absolute value of the difference between i and element number j (=|i−j|) is the same value or a close value, the sum of squares error E is also often the same value or a close value. In other words, it is difficult to reflect the asymmetry of penalties. As an example, case A where t=[0 1 0 0 0], y=[0.7 0.2 0.1 0 0] (that is, i=2, j=1), and t=[0 1 0 0 0] and y=[0.1 0.2 0.7 0 0] (that is, case B where i=2, j=3). When calculating the sum of squares error E for each case A and case B, |i−j| is equal to 1 in both cases, so the sum of squares error calculated by equation (1) is E for both cases A and B. =0.57.

Next, an example of a loss function that reflects the asymmetry of penalties will be explained. For example, a method of providing a coefficient _ak in the loss function equation as shown in equation (2) can be considered.

For example, equation (3) can be considered as a method for determining the coefficient a _k .

Here, i represents the element number with the largest value in the t array, that is, the true MCS index number. According to equation (3), the coefficient a _k when the element number k is larger than the element number i is a larger value than the coefficient a _k when the element number k is smaller than the element number i. Further, the coefficient a _k suddenly becomes larger as the element number k becomes larger than the element number i. On the other hand, when the element number k is less than or equal to the element number i, the coefficient a _k becomes smaller as the difference between the element number k and the element number i becomes larger, but the width of the decrease is small. That is, when element number k is larger than element number i, the penalty is larger than when element number k is less than or equal to element number i.

As an example, when the sum of squares error E is calculated using equation (2) for each of the above-mentioned cases A and B, in case A, E≈0.86, and in case B, E≈2.0. In this way, the sum of squares error E does not have the same value and reflects the asymmetry of the penalty.

Note that the method for determining the loss function in this embodiment is not limited to Equation (2) or Equation (3), but is limited to "the magnitude relationship between the estimated value j and the training data (= true value) i, and the asymmetry of the penalty. It is sufficient if the loss function is determined by considering the

According to the embodiment described above, MCS control can be performed predictively based on information obtained from camera images. Furthermore, compared to the conventional MCS control method that does not use camera images, it is possible to prevent a large amount of packet loss and improve throughput.

By using this embodiment, it is possible to improve the quality of a millimeter wave communication system in an environment where communication channel shielding frequently occurs, such as a station or commercial facility.

(Second embodiment)
In the first embodiment described above, an MCS index number representing a combination of a modulation method and a coding rate is estimated. In this embodiment, a new index number is assigned to each modulation method and coding rate in the MCS index table. The wireless communication system 1 uses an index number assigned to a modulation method or an index number assigned to a coding rate instead of the MCS index number. Thereby, the learning device 4 of the second embodiment directly estimates the modulation method and coding rate.

(Effects of this embodiment)
The effects of this embodiment compared with the conventional method calculated by simulation will be explained using FIGS. 5 and 6. As the conventional method, a conventional MCS control method that does not use camera images was used. Furthermore, we assumed that the environment in which the wireless communication equipment would be installed would be an environment where communication channels are frequently blocked, such as a station or commercial facility. FIG. 5 is a diagram showing the line disconnection probability when using the conventional method and the expected line disconnection probability when using the present embodiment. Further, FIG. 6 is a diagram showing a change in average throughput when using the conventional method and a change in average throughput expected when using this embodiment. From these figures, it can be seen that by using data such as camera images taken of the environment in which the communication device to which this embodiment is applied is installed, the line disconnection probability decreases as learning progresses, and is lower than that of the conventional method. Furthermore, as the learning progresses, the probability of line disconnection decreases, so it is expected that the average throughput will exceed that of conventional methods.

By using this embodiment, it is possible to improve the quality of a millimeter wave communication system in an environment where communication channel shielding frequently occurs, such as a station or commercial facility. Further, according to the present embodiment, learning data labeled with an index number representing a modulation method, a coding rate, or a combination of a modulation method and a coding rate is constructed in correspondence with a depth image, and A model representing correspondence with index numbers can be learned. Furthermore, even if the index number is incorrectly estimated, learning can be performed to prevent burst errors from occurring.

The learning device 4 may be implemented using a plurality of computer devices communicatively connected via a network. In this case, each functional unit included in the learning device 4 may be distributed and implemented in a plurality of computer devices. For example, the information management section 41, the learning data generation section 42, the machine learning section 43, and the estimation section 44 may be implemented in different computer devices, or some of them may be implemented in different computer devices. .

Further, the learning device 4 may be realized by a plurality of computer devices connected to a network. In this case, each of the functional units of the learning device 4 can be implemented by any one of the plurality of computer devices. For example, the information management section 41, the learning data generation section 42, the machine learning section 43, and the estimation section 44 may be implemented in different computer devices, or some of them may be implemented in different computer devices. . Furthermore, the same functional unit may be realized by a plurality of computer devices. Further, the BS 2 may have all or some of the functional units of the learning device 4.

When the functions of the learning device 4 in the embodiment described above are realized by a computer, a program for realizing this function is recorded on a computer-readable recording medium, and the program recorded on this recording medium is read into the computer system. This may also be achieved by setting and executing. It is also possible to provide this program through a network. Note that the "computer system" herein includes hardware such as an OS and peripheral devices. Furthermore, the term "computer-readable recording medium" refers to portable media such as flexible disks, magneto-optical disks, ROMs, and CD-ROMs, and storage devices such as hard disks built into computer systems. Furthermore, a "computer-readable recording medium" refers to a storage medium that dynamically stores a program for a short period of time, such as a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. It may also include a device that retains a program for a certain period of time, such as a volatile memory inside a computer system that is a server or client in that case. Further, the above-mentioned program may be one for realizing a part of the above-mentioned functions, or may be one that can realize the above-mentioned functions in combination with a program already recorded in the computer system.

An example of the hardware configuration of the learning device 4 will be described when the functions of the learning device 4 are realized by a computer. FIG. 7 is a device configuration diagram showing an example of the hardware configuration of the learning device 4. As shown in FIG. The learning device 4 includes a processor 71, a storage section 72, a communication interface 73, and a user interface 74.

The processor 71 is a central processing unit that performs calculations and control. Processor 71 is, for example, a CPU. The processor 71 reads the program from the storage unit 72 and executes it. The storage unit 72 further includes a work area when the processor 71 executes various programs. The communication interface 73 is communicably connected to other devices such as the BS 2 and the camera 3. The user interface 74 is an input device such as a keyboard, a pointing device (mouse, tablet, etc.), a button, a touch panel, or a display device such as a display. A human operation is input through the user interface 74 .

The functions of the information management section 41, learning data generation section 42, machine learning section 43, and estimation section 44 are realized by the processor 71 reading out and executing programs from the storage section 72. Note that all or part of these functions may be realized using hardware such as an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), or an FPGA (Field Programmable Gate Array).

According to the embodiment described above, the learning device includes a learning section. For example, the learning section corresponds to the machine learning section 43 in the embodiment. The learning unit acquires an image of the communication environment of the wireless communication system and the modulation method, coding rate, or combination of modulation method and coding rate that was used when wireless communication was successful in the wireless communication system. A model representing the correspondence between images and index values is learned based on data in which the images are associated with the index values. For example, the image is a camera image taken by camera 3. Also, for example, the model is a neural network. Further, for example, an MCS index representing a combination of a modulation method and a coding rate can be used as the index.

The learning unit learns the model so that the accuracy of the estimated value of the index estimated using the model is improved. The accuracy of the estimated value is unbalanced between cases where the estimated value of the index is larger than the true value and cases where the estimated value of the index is smaller than the true value.

The learning unit calculates the value of the loss function representing the accuracy of estimation using the estimated value of the index estimated using the model and the true value of the index, and estimates based on the calculated value of the loss function. Train the model to improve the accuracy of. There is an imbalance between the value calculated by the loss function when the estimated value of the index is large compared to the true value, and the value calculated by the loss function when the estimated value of the index is small compared to the true value. .

For example, the learning unit calculates the value of the loss function calculated using the estimated value of the index estimated using the model and the true value of the index, so that the communication speed corresponding to the estimated value of the index is the true value of the index. If the communication speed corresponding to the estimated value of the index is slower than the communication speed corresponding to the index, the communication speed is increased, and if the communication speed corresponding to the estimated value of the index is faster than the speed corresponding to the true value of the index, the communication speed is decreased. The learning unit learns the model to improve estimation accuracy based on the value of the loss function.

Also, for example, the value of the index is larger as the communication speed is faster, and the true value of the index is one consisting of elements t _k (k is an integer from 1 to K, inclusive) corresponding to each of the K values that the index can take. - The estimated value of the index is represented by an array t of hot expression, and the estimated value of the index is represented by an array y consisting of an element y _k that corresponds to each of the K values that the index can take and represents the probability of the corresponding value. . In this case, the loss function is a function that weights and adds the squares of the differences between element y _k and element t _k calculated for each element number k. The weight in the weighted addition increases as the element number k is larger than the element number corresponding to the true value. The element number corresponding to the true value is the element number i (i is any integer greater than or equal to 1 and less than or equal to K) set to 1 in the array t.

The learning device may further include an information management section and a learning data generation section. The information management unit stores a plurality of pieces of data in which images are associated with index values at the time the images were photographed. The learning data generation unit rewrites the index value associated with each of the multiple similar images to the value corresponding to the fastest communication speed among the index values associated with each of the multiple similar images. to generate learning data. The learning unit learns the model using the learning data generated by the learning data generation unit.

Although the embodiments of the present invention have been described above in detail with reference to the drawings, the specific configuration is not limited to these embodiments, and includes designs within the scope of the gist of the present invention.

The present invention is applicable to wireless communication systems in which communication channel shielding occurs.

DESCRIPTION OF SYMBOLS 1...Wireless communication system, 2...Base station, 3...Camera, 4...Learning device, 21...RF processing unit, 22...Communication control unit, 23...MCS control unit, 41...Information management unit, 42...Learning data generation unit , 43...Machine learning section, 44...Estimation section, 71...Processor, 72...Storage section, 73...Communication interface, 74...User interface

Claims (7)

An image taken of the communication environment of a wireless communication system and an index value representing the modulation method, coding rate, or combination of modulation method and coding rate that was used when wireless communication was successful in the wireless communication system. a learning step of learning a model expressing the correspondence between the image and the index value based on the data in which the image is associated with the index value;
has
In the learning step,
The value of the loss function calculated using the estimated value of the index estimated using the model and the true value of the index is determined such that the communication speed corresponding to the estimated value of the index corresponds to the true value of the index. If the communication speed is slower than the estimated value of the index, the communication speed corresponding to the estimated value of the index is made smaller than the speed corresponding to the true value of the index, and the estimated value is set based on the value of the loss function. A learning method for training said model to improve accuracy . The value of the index is larger as the communication speed is faster, and the true value of the index is one-, which is composed of elements t _k (k is an integer from 1 to K, inclusive) corresponding to each of the K values that the index can take. The estimated value of the index is represented by an array t of hot expression, and the estimated value of the index is represented by an array y consisting of elements y _k that correspond to each of the K possible values of the index and represent the probability of the corresponding value. In the case that
The loss function is a function that weights and adds the square of the difference between element y _k and element t _k calculated for each element number k,
The weight of the weighted addition is greater as the element number k is larger than the element number corresponding to the true value.
The learning method according to claim 1 . An image taken of the communication environment of a wireless communication system and an index value representing the modulation method, coding rate, or combination of modulation method and coding rate that was used when wireless communication was successful in the wireless communication system. a learning step of learning a model expressing the correspondence between the image and the index value based on the data in which the image is associated with the index value;
has
a storing step of storing a plurality of data in which the image is associated with the value of the index when the image was taken;
Learning data is obtained by rewriting the index value associated with each of the plurality of similar images to a value corresponding to the fastest communication speed among the index values associated with each of the plurality of similar images. a learning data generation step for generating
It further has
In the learning step, the learning method uses the learning data generated in the learning data generation step to learn the model. The model is a neural network,
The index is an MCS (Modulation and Coding Scheme) index representing a combination of a modulation scheme and a coding rate.
The learning method according to any one of claims 1 to 3 . An image taken of the communication environment of a wireless communication system and an index value representing the modulation method, coding rate, or combination of modulation method and coding rate that was used when wireless communication was successful in the wireless communication system. a learning unit that learns a model expressing the correspondence between the image and the index value based on the data in which the image is associated with the index value;
Equipped with
The learning department is
The value of the loss function calculated using the estimated value of the index estimated using the model and the true value of the index is determined such that the communication speed corresponding to the estimated value of the index corresponds to the true value of the index. If the communication speed is slower than the estimated value of the index, the communication speed corresponding to the estimated value of the index is made smaller than the speed corresponding to the true value of the index, and the estimated value is set based on the value of the loss function. A learning device that trains the model to improve accuracy . An image taken of the communication environment of a wireless communication system and an index value representing the modulation method, coding rate, or combination of modulation method and coding rate that was used when wireless communication was successful in the wireless communication system. a learning unit that learns a model expressing the correspondence between the image and the index value based on the data in which the image is associated with the index value;
Equipped with
a storage unit that stores a plurality of data associating the image with the value of the index when the image was taken;
Learning data is obtained by rewriting the index value associated with each of the plurality of similar images to a value corresponding to the fastest communication speed among the index values associated with each of the plurality of similar images. a learning data generation unit that generates
Furthermore,
The learning unit is a learning device that learns the model using the learning data generated by the learning data generation unit . computer,
A program for functioning as the learning device according to claim 5 or 6 .

Images