WO2022079832A1 - Communication information prediction device, communication information prediction method, and communication information prediction program - Google Patents

Abstract

The present invention comprises: an environmental information generation unit for generating environmental information on a device environment of at least one of a terminal device and a wireless communication destination device for the terminal device; a communication unit for generating communication information on wireless communication for the terminal device; a communication environment model generation unit for generating, by using input information including the environmental information, a communication environment model for commoditizing machine learning blocks formed by the same coefficients and structures and outputting communication information corresponding to a plurality of temporal conditions; and a model use unit for predicting communication information on the terminal device by using the generated communication environmental information. In this manner, parameters related to communication can be efficiently output under a plurality of temporal conditions.

Description

Communication information prediction device, communication information prediction method, and communication information prediction program

The present invention relates to a technique for predicting communication information of a wireless communication system corresponding to a plurality of time conditions.

In recent years, the realization of IoT (Internet of Things) in which various devices are connected to the Internet is progressing, and various devices such as automobiles, drones, and construction machinery vehicles are being connected wirelessly. It is also used for IEEE802.11 standard wireless LAN (Local Area Network), Bluetooth (registered trademark), LTE (LongTermEvolution) and 5G cellular communication, LPWA (Low Power Wide Area) communication for IoT, and vehicle communication. A wide variety of wireless communication standards such as ETC (Electronic Toll Collection System), VICS (Vehicle Information and Communication System (registered trademark)), and ARIB-STD-T109 have been developed, and various wireless communication services are widely used.

However, there is a problem that some wireless communication services cannot always meet the requirements for communication quality. In particular, when the terminal device or surrounding objects move, the directivity of the antenna and the propagation environment fluctuate, which greatly affects the communication quality. For example, in Non-Patent Document 1, a technique of generating a prediction model of wireless communication using distance information between a moving robot and a base station device and predicting communication quality is studied.

In a wireless communication system, radio wave propagation with a communication partner depends on the position of the terminal device equipped with a wireless communication function, the state of the terminal device such as posture and movement, and static or dynamic objects around the terminal device. The environment may change and communication quality may be affected, which may have a significant impact on the services and systems realized by wireless communication. For example, when a high frequency is used, the straightness of the radio wave is strong, and the communication quality is easily affected by changes in the radio wave propagation environment. Therefore, in order to manage the stable communication and communication quality of the terminal device to a higher degree, it is necessary to take measures against the fluctuation of the communication quality due to the terminal device itself and the surrounding environment of the terminal device.

As a countermeasure, the technique of modeling the relationship between environmental information and communication information using machine learning and predicting future communication quality is one of the useful means for maintaining good communication quality.

However, when outputting communication information such as future communication quality and communication-related events, future communication quality as reward conditions and maximization parameters, etc. based on environmental information and communication information, in what future Depending on how much time you are targeting, you will have to change the algorithms and models you use. In particular, when there are a plurality of time conditions, it is necessary to generate a plurality of models corresponding to each condition, and there is a problem that the load such as calculation cost and learning cost becomes large.

In view of the above problems, the present invention is a communication information prediction device and communication capable of generating a machine learning model that outputs communication-related parameters from input information including environmental information under a plurality of time conditions and predicting communication information. An object of the present invention is to provide an information prediction method and a communication information prediction program.

The present invention is a communication information prediction device that generates a communication environment model related to wireless communication of a mobile terminal device and predicts communication information of the terminal device, at least one of the terminal device and the wireless communication destination device of the terminal device. The environment information generation unit that generates the environment information related to the device environment, the communication unit that generates the communication information related to the wireless communication of the terminal device, and the input information including the environment information are used and configured with the same coefficient and structure. The terminal uses the communication environment model generator that generates a communication environment model that shares the machine learning block to be generated and outputs target information including communication information corresponding to a plurality of time conditions, and the generated communication environment model. It is characterized by having a model utilization unit that predicts communication information of the device.

Further, the present invention is a communication information prediction method for generating a communication environment model related to wireless communication of a mobile terminal device and predicting communication information of the terminal device, and is a wireless communication destination of the terminal device and the terminal device. The same coefficient is used by using the environment information generation process for generating environmental information related to at least one device environment of the device, the communication process for generating communication information related to the wireless communication of the terminal device, and the input information including the environment information. The communication environment model generation process that generates a communication environment model that outputs target information including communication information corresponding to multiple time conditions by sharing the machine learning block composed of the above and the structure, and the generated communication environment model. It is characterized in that it performs a model utilization process for predicting communication information of the terminal device.

Further, the present invention is a communication information prediction program for generating a communication environment model related to wireless communication of a mobile terminal device and causing a computer to execute a process of predicting communication information of the terminal device. Environmental information generation processing that generates environmental information related to the environment of at least one of the wireless communication destination devices of the terminal device, communication processing that generates communication information related to wireless communication of the terminal device, and input information including the environmental information. A communication environment model generation process that generates a communication environment model that outputs target information including communication information corresponding to multiple time conditions by sharing a machine learning block composed of the same coefficient and structure using It is characterized in that a computer executes a model utilization process for predicting communication information of the terminal device using the communication environment model.

The communication information prediction device, the communication information prediction method, and the communication information prediction program according to the present invention generate a machine learning model that outputs communication parameters from input information including environmental information under a plurality of time conditions, and communication information. Can be predicted. In particular, by using a common machine learning block, the learning cost and the model usage cost can be reduced.

It is a figure which shows the configuration example of a wireless communication system. It is a figure which shows the configuration example in the case of training a communication environment model generation part. It is a figure which shows an example of the 1st communication environment model which concerns on this embodiment. It is a figure which shows an example of the 2nd communication environment model which concerns on this embodiment. It is a figure which shows an example of the autonomous mobile robot. It is a figure which shows the indoor map and the position of a goal which an autonomous mobile robot recognizes. It is a figure which shows the specific example of the deep neural network of the communication environment model of 1st configuration. It is a figure which shows the specific example of the deep neural network of the communication environment model of the 2nd configuration. It is a figure which shows the relationship between the input and output of RNN. It is a figure which shows an example of the functional block of the communication environment model of 1st configuration. It is a figure which shows an example of the functional block of the communication environment model of the 2nd configuration. It is a figure which shows the other structural example of the 4th machine learning block. It is a figure which shows an example of the communication environment model of the comparative example of the demonstration experiment. It is a figure which shows the result of the performance evaluation in the demonstration experiment.

Hereinafter, embodiments of the communication information prediction device, the communication information prediction method, and the communication information prediction program according to the present invention will be described with reference to the drawings. The embodiments described below are examples of the present invention, and the present invention is not limited to the following embodiments.

FIG. 1 shows a configuration example of the wireless communication system 100. The wireless communication system 100 is composed of M units (M is a positive integer) of terminal devices 102 (1) to 102 (M) that perform wireless communication with the base station device 101 and the base station device 101. Here, when the description common to the M terminals 102 (1) to 102 (M) is given, the terminal device 102 is described by omitting the (number) at the end of the code. The same applies to a plurality of the same blocks constituting the terminal device 102.

In the example of FIG. 1, all of the base station device 101 and the M terminal devices 102 have a function as the communication information prediction device according to the present invention, and generate a communication environment model for predicting communication information. Communication information can be predicted.

In the present embodiment, the configuration of connecting to the network via the base station device 101 will be described, but a communication device capable of connecting to the network may be provided separately from the base station device 101. Then, the terminal device 102 is a device (access point including the base station device 101 or a monitoring camera on a peripheral road) equipped with a camera or a sensor connected to the network via the base station device 101 or a communication device connectable to the network. Non-communication information such as camera information and sensor information can be collected as environmental information. Further, the terminal device 102 may directly communicate with another terminal device 102 to collect camera information and sensor information from the camera / sensor mounted on the terminal device 102. For example, when the terminal device 102 is a communication device mounted on an automobile, camera information and sensor information (including position information) are collected by vehicle-to-vehicle communication (vehicle-to-vehicle communication). In this way, environmental information for generating a model of wireless communication between the target base station device 101 and the terminal device 102 by machine learning is collected. Then, the communication information between the base station device 101 and the terminal device 102 is predicted using the generated communication environment model.

(Configuration example of base station device 101)
In FIG. 1, the base station device 101 has a function of connecting to a network and performs wireless communication with M terminal devices 102. The base station apparatus 101 includes a NW unit 201, a communication unit 202 (1) to a communication unit 202 (N), a communication environment model generation unit 203, an environment information generation unit 204, and a model utilization unit 205.

The NW unit 201 is an interface for connecting to an external network. The base station device 101 can acquire camera information (photographed image, etc.) and sensor information (three-dimensional laser, etc.) from a device such as a camera or sensor connected to an external network via the NW unit 201. This information can also be acquired by the terminal device 102 connected to the base station device 101.

The communication unit 202 has N numbers (N is a positive integer) from the communication unit 202 (1) to the communication unit 202 (N), and can perform wireless communication with a plurality of terminal devices 102 (N). Communication processing). Further, the communication unit 202 generates communication information related to wireless communication and outputs it to the communication environment model generation unit 203 and the model utilization unit 205. Communication information includes, for example, received signal power, signal-to-noise power ratio (SNR), signal-to-noise power ratio (SINR), RSSI (ReceivedSignalStrengthIndication), RSRQ (ReceivedSignalReferenceQuality), packet error rate, and arrival. The number of bits, the number of reached bits per unit time, MCS (Modulation and Coding Scheme index), the number of retransmissions, the delay time, the error correction method, the frequency of the communication system, the frequency conditions such as the bandwidth of the resource to be used, and the like. Further, it may be the differential information of these values, an index calculated from these values using a calculation formula, a system setting item that affects these indexes, and the like.

The communication environment model generation unit 203 includes at least one of the environment information generated by the environment information generation unit 204, the communication unit 202 (1) to the communication unit 202 (N), and the communication unit 301 (1) to the communication unit 301 (M). Using the communication information generated by one, a communication environment model for predicting the communication information is generated (communication environment model generation process). The method of generating the communication environment model will be described later.

The environmental information generation unit 204 generates environmental information related to the device environment (environmental information generation processing). Environmental information includes camera information and sensor information (of an object) of a device that can be acquired via the position / attitude / state / movement / control command / terminal device 102 of the terminal device 102 and the base station device 101 or the base station device 101 or a network. Includes at least one piece of information such as (including various information such as presence / absence, size, material) / operation of structure / action strategy of terminal device 102 and network.

Here, the control command corresponds to, for example, when the terminal device 102 is an autonomous mobile robot, a tire rotation command in the front direction of the robot, a rotation command of the robot, and the like. Further, the action strategies of the terminal device 102 and the network include, for example, the planned movement route of the terminal device 102 and the robot, the power mode of the terminal device 102, the number of antennas used by the terminal device 102, the QoS setting to the terminal device 102, and the network. Route settings, network congestion status, network routing rules, application settings, etc.

Furthermore, the environmental information is target information in the communication environment model, such as owner information and type information (mobile terminals such as smartphones, automobiles, robots, drones, etc.), store occupant information, automobile movement history, etc. of the terminal device 102. It may be information including all non-communication information that may be used to predict. The more environmental information there is, the better the prediction accuracy of the communication environment model may be.

The model utilization unit 205 outputs target information or strategic information that maximizes the reward by using the communication environment model generated by the communication environment model generation unit 203 (model utilization processing). For example, when the terminal device 102 is an autonomous mobile robot, it is possible to set a reward for maximizing the received power and output X-coordinate speed, Y-coordinate speed, rotation command, etc. that increase the reward as strategic information. can. The operation of the model utilization unit 205 will be described later.

In this way, when the base station device 101 functions as a communication information prediction device, the communication environment model generation unit 203 includes the environmental information obtained from the environment information generation unit 204, the communication information obtained from the communication unit 202, and the communication information. To generate a communication environment model using. Then, the model utilization unit 205 can predict the communication information by using the communication environment model generated by the communication environment model generation unit 203.

(Configuration example of terminal device 102)
The terminal device 102 (1) to the terminal device 102 (M) each perform wireless communication with the base station device 101.

The terminal device 102 has a communication unit 301, a communication environment model generation unit 302, an environment information generation unit 303, and a model utilization unit 304.

The communication unit 301 performs wireless communication with the communication unit 202 of the base station device 101 or another terminal device 102 (communication processing). Further, the communication unit 301 generates communication information related to wireless communication and outputs it to the communication environment model generation unit 302 and the model utilization unit 304, similarly to the communication unit 202 of the base station apparatus 101.

Similar to the communication environment model generation unit 203 of the base station apparatus 101, the communication environment model generation unit 302 communicates using the environment information obtained from the environment information generation unit 303 and the communication information obtained from the communication unit 301. Generate an environment model (communication environment model generation process).

The environmental information generation unit 303 generates environmental information (environmental information generation processing) in the same manner as the environmental information generation unit 204 of the base station apparatus 101. The method of generating the communication environment model will be described later.

Similar to the model utilization unit 205 of the base station apparatus 101, the model utilization unit 304 uses the communication environment model generated by the communication environment model generation unit 302 to output target information or maximize the reward. Output information (model usage process).

In this way, when the terminal device 102 functions as a communication information prediction device, the communication environment model generation unit 302 can obtain the environmental information obtained from the environment information generation unit 303 and the communication information obtained from the communication unit 301. Use to generate a communication environment model. Then, the model utilization unit 304 can predict the communication information by using the communication environment model generated by the communication environment model generation unit 302.

Here, in the above description, the base station device 101 and the terminal device 102 have been described separately, but the functions may be shared between the base station device 101 and the terminal device 102. For example, the communication environment model generated by the communication environment model generation unit 203 may be used by the model utilization unit 304 of each terminal device 102.

(Communication environment model generator)
Next, the communication environment model generation unit 203 of the base station device 101 and the communication environment model generation unit 302 of the terminal device 102 will be described in detail. Although the description is given here as the communication environment model generation unit 203 of the base station device 101, the basic operation is the same for the communication environment model generation unit 302 of the terminal device 102.

Since the communication environment model generation unit 203 of the base station device 101 can collect communication information and environment information in wireless communication with a plurality of terminal devices 102 and use a lot of training data, the communication of the terminal device 102 can be used. It is preferable to the environment model generator 302. However, it is possible to adjust by techniques such as transfer learning and fine tuning so that the model is optimized for each terminal device 102.

The communication environment model is a model generated so as to output environmental information or target information at the same time or future time by using environmental information and communication information as input information. The communication environment model may be generated so that the output value is highly accurate, or may be generated by reinforcement learning so as to maximize a certain parameter different from the output value.

For example, when the received power of the future time is output by the communication environment model, the training data is formed by using the received power of the future time as the parameter of the target information, and the predicted value of the received power that is the output of the communication environment model is used. The communication environment model is generated so that the error from the actual received power is small. Alternatively, when the terminal device 102 is an autonomous mobile robot, the received power is maximized by enhanced learning in order to output the X-coordinate speed, Y-coordinate position speed, and rotation command of the robot so as to maximize the received power. A communication environment model is generated so as to set a reward for the robot and output an X-coordinate velocity, a Y-coordinate velocity, and a rotation command so as to increase the reward.

The communication environment model can be trained efficiently by using a common structure and output, and it is thought that it is easy to optimize the coefficients of the model.

FIG. 2 shows a configuration example when training the communication environment model generation unit 203. Although the description is given here as the communication environment model generation unit 203 of the base station device 101, the same applies to the communication environment model generation unit 302 of the terminal device 102.

In FIG. 2, the communication environment model generation unit 203 (or communication environment model generation unit 302) communicates using environment information or training data composed of input information consisting of environment information and communication information and target information. Generate the environment model 401. At the time of training, the communication environment model 401 generated to output appropriate target information corresponding to the input information, at the time of use, obtains the target information of the future time from the environmental information or the input information consisting of the environmental information and the communication information. Output.

Here, the target information is communication information or a parameter related to control that maximizes the parameter of the communication information. The control parameters include the movement of the terminal device 102, the movement of the wireless communication destination device such as the base station device 101 or another terminal device 102, and the mode / communication destination / communication path / in the OSI reference model from the physical layer to the application layer. Control of communication method, etc., control of structure / metamaterial / dielectric position / movement / setting, etc. that affect the radio wave propagation environment.

FIG. 3 shows an example of the first communication environment model 401 (1) according to the present embodiment. The first communication environment model 401 (1) is composed of a first machine learning block 501 and second machine learning blocks 502 (1) to (K). K is the number of corresponding times of the output target information, and the common first machine learning block 501 is used to generate the target information corresponding to different times. In FIG. 3, the first machine learning block 501 is configured using an arbitrary machine learning algorithm such as a neural network, a decision tree, or a random forest.

Here, the output information of the first machine learning block 501 is defined as the feature space information. The feature space information is environment information or information generated by performing some kind of signal processing on the environment information, and is information that includes information about the environment information or can be extracted from the environment information. For example, information effective for the second machine learning block to output target information corresponding to different times is output from the first machine learning block 501 as feature space information. It should be noted that the first machine learning block 501 can also be constructed by a recurrent neural network (RNN) in order to output for different times.

In FIG. 3, the feature space information output by the first machine learning block 501 is input to the second machine learning block 502 that outputs the target information corresponding to each of the plurality of time conditions.

Here, the second machine learning block 502 is an individual block represented by a different coefficient corresponding to each of the plurality of time conditions, and a plurality of blocks having the same structure are provided for each time condition. On the other hand, the first machine learning block 501 having the same coefficient and the same structure is shared, and the output of the first machine learning block 501 is commonly used for the second machine learning block 502. In the example of FIG. 3, as target information of a plurality of future time conditions, time t _{F, 1} target information 503 (1), time t _{F, 2} target information 503 (2), ..., Time t _{F , K} target information 503 (K) target information 503 for K time conditions (K is a positive integer) is output.

By sharing the first machine learning block 501 in this way, it is not necessary to individually provide the first machine learning block 501 corresponding to different time conditions, and efficient learning becomes possible.

FIG. 4 shows an example of the second communication environment model 401 (2) according to the present embodiment. In FIG. 4, in the communication environment model 401 (2), when the input information is input to the third machine learning block 601, the feature space information 602 corresponding to a plurality of time conditions is output.

Here, the first machine learning block 601 has been described with reference to FIG. 3 in that the third machine learning block 601 outputs a plurality of feature space information 602 corresponding to a plurality of time conditions output by the fourth machine learning block 603. Different from machine learning block 501. In the example of FIG. 4, the third machine learning block 601 has the feature space information 602 (1) at time t _{F, 1} , the feature space information 602 (2) at time t F, 2, ..., Time t _{F. , K} feature space information 602 (K) K feature space information 602 is output.

The feature space information 602 corresponding to different times output from the third machine learning block 601 is input to the fourth machine learning block 603 shared by the same structure. The fourth machine learning block 603 is a machine learning block that uses feature space information of a certain time condition and outputs target information of the same time condition as the time. For example, in FIG. 4, the target information 604 (1) under the same time condition is output from the feature space information 602 (1) at time t _{F, 1} . Similarly, from the feature space information 602 (2) to the time t _{F, 2} target information 604 (2), ..., From the feature space information 602 (L) to the time t _{F, K} target information 604 (K) K Target information 604 for each time condition is output.

As described above, the communication environment model 401 (2) of the second configuration shares the fourth machine learning block 603 without the need to train independent machine learning blocks for each of the plurality of time conditions. Therefore, it is possible to efficiently generate a communication environment model.

The third machine learning block 601 is configured by using the first machine learning block 501 and the second machine learning block 502 described with reference to FIG. 3, so as to generate feature space information for a plurality of time conditions. May be good.

[Demonstration experiment]
The experiments conducted to demonstrate the effects of the embodiments and the results thereof will be described.

The mobile terminal device 102 used in the experiment uses LIDAR ("Light Detection and Ringing" or "Laser Imaging Detection and Ringing"), which is one of the remote sensing technologies using pulsed laser irradiation (scattering). It is an on-board autonomous mobile robot. The autonomous mobile robot can collect its own position information, orientation information, odometry information, and control command information as terminal information by using LIDAR.

In addition, wireless communication uses a wireless LAN of the IEEE802.11ac standard, and the amount of bits communicated in 0.2 seconds is measured as a throughput to evaluate the performance.

FIG. 5 shows an example of an autonomous mobile robot. In FIG. 5, the horizontal axis represents a two-dimensional plane of x-axis [m] and the vertical axis represents y-axis [m], and the position of the autonomous mobile robot at time τ is (x [τ], y [τ]). Is. The self-position information is calculated by the algorithm of AMCL (Adaptive Monte Carlo Localization) of Robot OS from the odometry information and the measurement result of LIDAR. The orientation is represented by {kz [τ], w [τ]} = {sin (θ [τ] / 2), cos (θ [τ] / 2)} by the quaternion used in the coordinate system in the robot field. To. In addition, odometry information such as speed and rotation speed with respect to the front direction of the robot detected from the rotation of the robot tire, tire rotation command corresponding to the front direction of the robot, control command information such as a command corresponding to the rotation of the robot, etc. As the feature amount (parameter), odometry (Vx [τ], _Vθ [τ]) and control command information (Ox [τ], _Oθ [τ]) can be used. In the demonstration experiment using the autonomous mobile robot described later, (x [τ], y [τ], kz [τ], w [τ], Vx [τ], V _θ [τ], Ox [τ], O Of the total of 8 features of _θ [τ]), all the features (8 features) are used as input information, and the features other than the control command information (6 features) are used as input information. When to use and when to do.

Furthermore, communication information uses RSSI and throughput, RSSI is acquired every 10 ms, the median RSSI in the 0.2 second time slot is γ _ave [τ], the dispersion value is γ _std [τ], and the throughput. Let be C [τ].

Here, the environmental information in the demonstration experiment is the position (X-axis, Y-axis), orientation, odometry (self-position estimation) (velocity on X-axis, velocity on Y-axis, rotational velocity) and control obtained from the robot. Command information (tire rotation command in the front direction of the robot, rotation command of the robot), and past throughput (from the present to the past 2 seconds) are used.

FIG. 6 shows an indoor map recognized by an autonomous mobile robot and the position of a goal. As shown in FIG. 6, the robot selects one of the 113 goals set and continues autonomous movement. Then, when the robot enters 30 cm around the selected goal, the next goal is reset and the robot is designed to move toward the next goal. In the demonstration experiment, the communication environment model 401 of the first configuration described in FIG. 3 for future communication information from the data set obtained over 60 hours based on the result of the random walk of the robot by automatic operation ( The predicted results are evaluated using the communication environment model 401 (2) according to the second configuration described in 1) and FIG.

FIG. 7 shows a specific example of the deep neural network of the communication environment model 401 (1) of the first configuration. In FIG. 7, the communication environment model 401 (1) having the first configuration has a first machine learning block 501 and a second machine learning block 502 as described with reference to FIG. The first machine learning block 501 is composed of an RNN, and the second machine learning block 502 is composed of three fully connected layers.

In FIG. 7, the horizontal axis represents the time axis, and the input information of the first machine learning block 501 is represented by Ω [ _ti ] corresponding to a plurality of times. Then, with the current time as t ₀ , the past 1 second is every Δt = 0.2 seconds, t ₀ , t _-1 = t 0-0.2 seconds, t _{-2 =} t _0-0.4 seconds, t. Ten timing parameters of _-3 = t _0-0.6 seconds, t _{-4 =} t _0-0.8 seconds, ..., T- ₉ = t 0-1.8 seconds are input.

Here, Ω [t] is changed to {X [t], Y [t], kz [t], w [t], Vx [t], V _θ [t], Ox [t], O _θ [t]. }, In FIG. 7, the number of input information parameters (number of input elements: Input Min) is 8, and the number of inputs _in the time direction is 10, so the number of parameters used is 8 ×. 10 = 80. When expressed as discrete values, the timing is represented by time t ₀ , t _-1 , t _-2 , ..., T- ₉ , respectively.

In FIG. 7, the dotted white-painted parallelogram block represents the input information parameter at each timing, and the gray-painted parallelogram block represents a layer such as a hidden layer or a fully connected layer of the neural network. In the example of FIG. 7, the second machine learning block 502 is a target of time series data corresponding to future time conditions (five times of time t _F-4 , t _F-3 , ..., T _F ). Predicting information (throughput in FIG. 7), Output: C [t _F-4 ], C [t _F-3 ], C [t _F-2 ], C [t _F-1 ], C [t _F ] ] Is output.

GRU (Gated Recurrent Unit) is used for the RNN of the first machine learning block 501, the hidden layer is 1, and the dimension of the hidden layer is 35.

In the RNN, each layer is connected from the old layer to the new layer in time, so that the temporal relationship is maintained. In the example of FIG. 7, from the layer of time t- ₉ to the layer of time t- ₈ , ..., from the layer of time t _-4 to the layer of time t _-3 , from the layer of time t _-3 to the layer of time t ₋ Each layer is coupled to the _second layer, from the time t _-2 layer to the time t _-1 layer, and from the time t _-1 layer to the time t ₀ layer. In this way, in the RNN, processing such as weight multiplication and bias addition is performed in order from the past information.

The fully connected layer of the second machine learning block 502 is composed of two layers of the output 35 at the input 35 and one layer of the output 1 at the input 35, for a total of three layers.

In FIG. 7, if the set of input information at time t _i is Ω [ti _] , the input information from time t- ₉ to t ₀ is processed by RNN, and the input information from the final layer at time t ₀ is from the 35th dimension. The output is commonly used as the input for the second machine learning block 502.

Although the RNN block can be output from a layer other than the final layer, in the example of FIG. 7, the RNN block is output from the final layer of the RNN block, and a plurality of fully coupled devices corresponding to each of the plurality of times of the second machine learning block 502 are fully connected. It is configured to be input in common to layers.

Here, the five fully connected layers of the second machine learning block 502 corresponding to the five times from time t _F-4 to t _F input the same information output from the first machine learning block 501. .. Then, processing such as weight multiplication and bias addition is performed in each fully connected layer, and the throughputs C [t _F-4 ], C [t _F-3 ], and C [t corresponding to the five times are performed. _F-2 ], C [t _F-1 ], and C [t _F ] are output, respectively.

Although omitted in FIG. 7, an activation layer by ReLU (Rectified Linear Unit) is used between the fully connected layers. At the time of training, a dropout layer may be inserted to randomly eliminate the input / output between the neural networks. Furthermore, both the input information and the target information are methods used for preprocessing and dimensional compression (PCA: Principal Component Analysis, LDA: Linear Discriminant Analysis) so that the variance and range are changed from -1 to +1 or 0 to 1. By doing such as), it is possible to generate a model efficiently.

Here, the prediction accuracy of the demonstration experiment was evaluated by the R2 score of the equation (1).

In addition, C _i is the measured value of the throughput, (^ C _i ) is the predicted value by machine learning, n is the number of samples, i is the sample number, and _Ave is the predicted value of the throughput.

FIG. 8 shows a specific example of the deep neural network of the communication environment model 401 (2) of the second configuration. In FIG. 8, the communication environment model 401 (2) having the second configuration has a third machine learning block 601 and a fourth machine learning block 603, as described with reference to FIG.

In FIG. 8, the third machine learning block 601 is composed of an RNN, three fully connected layers, and a second RNN, and the fourth machine learning block 603 is composed of three fully connected layers. ..

In FIG. 8, the horizontal axis represents the time axis, and the input information of the third machine learning block 601 is represented by Ω [ _ti ] corresponding to a plurality of times. Then, with the current time as t ₀ , the past 1 second is every Δt = 0.2 seconds, t ₀ , t _-1 = t 0-0.2 seconds, t _{-2 =} t _0-0.4 seconds, t. Ten timing parameters of _-3 = t _0-0.6 seconds, t _{-4 =} t _0-0.8 seconds, ..., T- ₉ = t 0-1.8 seconds are input.

Here, the RNN of the third machine learning block 601 is configured in the same manner as the first machine learning block 501 of FIG. 7. That is, with the set of input information at time ti as Ω [ti _] , processing such as weight multiplication and bias addition is performed in order from the past information in the RNN, and the input from time t- ₉ to t ₀ is _performed . The output from the 35th dimension of the final layer at time t ₀ when the information was processed is commonly used in the next three fully coupled layers. That is, the information output from the RNN is commonly input to the eight fully connected layers corresponding to the eight times from time t _F-7 to t _F , respectively.

Then, after processing such as weight multiplication and bias addition is performed in the eight fully connected layers, the _M'out of the intermediate output 1 is output and input to the second RNN. The second RNN takes the _M'out of the intermediate output 1 as an input, and the feature spatial information Θ [t _F-4 ] to Θ [t _F ] corresponding to the five future times t _F-4 to t _F to be targeted. ] Is output as intermediate output 2.

The fourth machine learning block 603 converts and outputs the target information corresponding to the same time using the feature space information Θ [t _F-4 ] to Θ [t _F ] in the time t _F-4 to t _F. do. The target information is, for example, the throughputs C [t _F-4 ], C [t _F-3 ], C [t _F-2 ], C [t _F-1 ], C [t _F ] corresponding to each of the five times. Is.

In the configuration of FIG. 8, since the third machine learning block 601 can generate the feature space information Θ corresponding to an arbitrary time, the fourth machine learning block 603 is configured with the same weight, bias, and other coefficients. The desired target information can be predicted using the machine learning block. That is, the fourth machine learning block 603 only needs to model the relationship between the feature space information and the target information in the same future time. The fourth machine learning block 603 can efficiently perform training without depending on the condition of how much future time target information is predicted for the time t ₀ when the information is acquired. can. Thus, in this embodiment, a common model can be used to predict target information for a plurality of future times.

Here, the third machine learning block 601 of FIG. 8 has a configuration in which the first machine learning block 501 and the second machine learning block 502 of the first configuration of FIG. 7 are combined. The difference from FIG. 7 is that the environmental information from the time t _F-4 to t _F is output as the feature space information.

During training, the values of weights and biases in the third machine learning block 601 by backpropagation from the collected data set based on the error between the collected data set and the environmental information at t _F-4 to t _F in the future time. To update. In this way, since the time-series data of the environmental information of the future time can be obtained, the feature space information Θ having a high relationship with the target can be generated by the second RNN.

In the first machine learning block 501 described with reference to FIG. 7, the final layer of the RNN was used as a common input to the fully connected layer. On the other hand, the fourth machine learning block 603 of FIG. 8 outputs the target information corresponding to different times by using the intermediate output 2 of the intermediate layer of the second RNN of the third machine learning block 601. can do.

FIG. 9 is a diagram showing the relationship between the input and output of the RNN of FIG. In FIG. 9, the RNN uses the input information Ω [ti] of L time slots (L is a positive integer) from the time t _{Z- (L-1)} to t _Z , and the time t _{M- (L)} . It is shown that the output information Θ [ti _] corresponding to L times from _-1) to t _M can be obtained. In this way, the RNN operates so as to gradually propagate past information toward the future, but on the way, it is possible to output an output corresponding to an arbitrary time up to that point. For example, in FIG. 9, in the state of Z ₌ 0, it is possible to train to output the feature space information Θ [ti] corresponding to t _M from the future time t _{M − (L-1)} ahead of the M time slot. .. Here, when M = Z, Θ generated by using the input information up to Ω [Mi] and Ω [j] (j≤Mi) with respect to the target information at time _tMi . It can be expected that [tM _-i ] has a high correlation. On the other hand, when outputting the feature space information Θ such that M> Z, Ω [t _M ], which is the result including all up to Ω [t _Z ], is the time t _{M i} . It can be expected to have a high correlation with the target information. This is because the time difference t _M-i- t _Z is smaller than the time difference t _M-i- t _Z-i .

The usage of the output of the RNN will be described by taking two examples, the RNN of FIG. 8 and the second RNN. The first RNN in FIG. 8 generates feature space information Θ corresponding to future time with respect to the input information. In FIG. 9, M> Z. In this case, since the time difference t _F- t ₀ is smaller than the time difference t _{F-i- t} -i, the last information is output, so that future input information can be included as much as possible. .. In the example of FIG. 9, only the last output information Θ [t _M ] is output, and up to new information (time t _Z ) can be included in the output as much as possible in terms of time, and the prediction accuracy is improved.

On the other hand, in the second RNN of FIG. 8, the input time and the output time are the same, and M = Z in FIG. In this case, even under the condition that the information up to _tFi is used to output _tFi , the time difference between the corresponding times of the input and the output becomes 0, and the target information can be expected to be highly accurate. Output information of multiple times on the way. Here, in the example of FIG. 8, since it is desired to obtain the output information of the five time series in the time series, the information of the hidden layer of the latter five hours out of the eight hidden layers of the second RNN is output as the intermediate output 2. Has been done.

In this way, in the communication environment model of the first configuration and the second configuration according to the present embodiment, the target information of an arbitrary time and an arbitrary time width can be obtained by the machine learning block using the RNN. ..

FIG. 10 shows an example of a functional block of the communication environment model 401 (1) having the first configuration of FIG. 7.

In FIG. 10, the output of the final layer of the RNN of the first machine learning block is input in common to different fully connected layers, and the target information corresponding to a plurality of different times is output respectively.

In FIG. 10, the RNN of the first machine learning block 501 uses GRU, the hidden layer is 1, and the dimension of the hidden layer is 35. Further, the second machine learning block 502 is configured for each of a plurality of times by three fully connected layers having two layers of output 35 at input 35 and one layer of output 1 at input 35. Will be done.

Input information of environment information (8 parameters) and communication information (2 parameters) is input to the RNN of the first machine learning block 501. The output of the RNN is used as a common input for a second machine learning block 502 consisting of three fully coupled layers corresponding to multiple times, where the second machine learning block 502 is in the future time _{tF- (L)} . _-1) Outputs target information corresponding to a plurality of times of, t _{F- (L-2)} , ..., T _F , respectively.

FIG. 11 shows an example of the functional block of the communication environment model 401 (2) having the second configuration of FIG.

In FIG. 11, a common fully connected layer of the fourth machine learning block 603 is used for outputs corresponding to a plurality of different times output from the third machine learning block 601. Here, as the fourth machine learning block, there are a plurality of block lines, but the coefficients such as weights and biases of the neural networks existing in parallel have the same configuration, and the feature space information Θ corresponding to a certain time. Is converted into the target information corresponding to the same time.

In FIG. 11, the RNN of the first machine learning block 501 uses GRU, the hidden layer is 1, and the dimension of the hidden layer is 35. Further, the second machine learning block 502 is configured for each of a plurality of times by three fully connected layers having two layers of output 35 at input 35 and one layer of output 1 at input 35. And output the intermediate information. The intermediate information includes environment information and communication information corresponding to the intermediate output 1 in FIG. 8, and the number of samples in the time direction is L. The second RNN is composed of a hidden layer 1 and a dimension 35 of the hidden layer, and outputs information for each of L times out of a plurality of times (corresponding to the intermediate output 2 in FIG. 8). Then, the output of the second RNN is used as a common input of the fourth machine learning block 603 composed of three fully connected layers, and the fourth machine learning block 603 is used in the future time _{tF- (L-)} . ₁₎ , t _{F- (L-2)} , ..., As target information corresponding to a plurality of times of t _F , throughput C [t _{F- (L-1)} ], C [t _{F- (L-} ). ₂₎ ], ..., C [t _F ] are output respectively.

FIG. 12 shows another configuration example of the fourth machine learning block 603 of FIG. In FIG. 11, after all the information of the environment information and the communication information is once generated as the intermediate information, the information is input to the fourth machine learning block 603. On the other hand, in FIG. 12, only parameters related to environmental information such as the position of the robot are first generated as intermediate information, and the second RNN is the target time interval (t _{F- (L-1)} to t _F ). The feature space information Θ [t _{F- (L-1)} ] corresponding to Θ [t _F ] is generated.

Then, the information is copied into two, and one of them is input to a three-layer coupling layer (input 35 has two layers of output 35 and input 35 has one layer of output 3), and a second layer is input every time. Communication information (parameters related to throughput and RSSI, in this case, γ _ave [t], γ _std [t], and C [t]) is output as intermediate information. The coefficients in the neural network are updated by back propagation so that the accuracy of the second intermediate information is improved.

Further, the second intermediate information and the other feature space information Θ copied earlier are output by three fully connected layers (input 38 at output 35, input 35 at output 35, and input 35). It is input to the layer 1), and the target communication information (here, C [t _{F- (L-1)} ] to C [t _F ]) is output.

FIG. 13 shows an example of a communication environment model of a comparative example of the demonstration experiment. A comparative example is a conventional communication environment model, which is a deep learning block that predicts future throughput by using a common machine learning block. FIG. 13 shows an example of a communication environment model configured by using an RNN and three fully connected layers.

In FIG. 13, the RNN uses a GRU, the hidden layer is 1, and the dimension of the hidden layer is 35. Further, the fully connected layer is three layers, two layers of input 35 and output 35, and one layer of input 35 and output 5. An activated layer by ReLU is used between the fully bonded layers.

As described above, in the configuration of the comparative example, the output of the final layer of the RNN block is given to the fully connected layer, and the final layer outputs the target information corresponding to a single time. Therefore, in the comparative example, there is a problem that it is difficult to obtain a flexible output for an arbitrary time position corresponding to a plurality of time conditions such as how the target information changes with time.

Next, the results of demonstration experiments in each configuration of the communication environment model 401 (1) of the first configuration, the communication environment model 401 (2) of the second configuration, and the communication environment model of the comparative example according to the present embodiment. Will be explained. Here, the results of performance evaluation when predicting the future throughput by the R2 score are compared.

As the experimental data, a total of 96 hours of data set collected from the autonomous mobile robot in the above-mentioned indoor experimental environment was used. Using a 20 MHz channel in the 5.6 GHz band, X [t], Y [t], kz [t], w [t], Vx [t], V _θ [t], Ox [t], O _θ [t] ], Γ _ave [t], γ _std [t], and C [t] are evaluated using the collected data set, and 10 hours of the data are used as test data and evaluated by the R2 score. Was done.

Here, the input information used in the demonstration experiment is in the case of six parameters {X [t], Y [t], kz [t], w [t], Vx [t], V _θ [t]}. And {X [t], Y [t], kz [t], w [t], Vx [t], V _θ [t], Ox [t], O _θ [t]} There are two cases. In any case, as the intermediate information shown in FIG. 12, the environmental information {X [t], Y [t], kz [t], w [t], Vx [t], V _θ [t]} 6 As one parameter and the second intermediate information, three parameters of communication information {γ _ave [t], γ _std [t], C [t]} are output, respectively.

Another effect of this embodiment is that the input information is {X [t], Y [t], kz [t], w [t], Vx [t], V _θ [t]}. In the case of parameters, or {X [t], Y [t], kz [t], w [t], Vx [t], V _θ [t], Ox [t], O _θ [t]} 8 In the case of two parameters, since the intermediate information is the same, the coefficients of the neural network shown in FIG. 12 can be shared. That is, by defining the intermediate information in advance, it is possible to output time-series data of a plurality of time conditions corresponding to various input information.

FIG. 14 shows the results of performance evaluation in the demonstration experiment. From the data set for 96 hours of the entire demonstration experiment, 80 hours was used for training, 6 hours was used for accuracy verification data (Validation), and 10 hours was used for evaluation as test data. The training was performed with the optimization algorithm (Optimizer) set to Adam and the learning rate set to 2.0 × 10-4, and training was performed until the MSE (Mean Square Error) for the accuracy verification data was minimized to generate a communication environment model. ..

In addition, the target throughput is standardized at 150 Mbps, and for each time of 5 time slots (t ₁₀ , t ₉ , t ₈ , t ₇ , t ₆ ) after 1.0 to 2.0 seconds. The R2 score was evaluated against the average of the throughputs C [t ₁₀ ], C [t ₉ ], C [t ₈ ], C [t ₇ ], and C [t ₆ ].

6 types of input information {X [t], Y [t], kz [t], w [t], Vx [t], V _θ [t]}, and 8 types of input information {X [t], Generate models for Y [t], kz [t], w [t], Vx [t], V _θ [t], Ox [t], O _θ [t]} and apply them to the test data. bottom. The R2 score is a predicted value in the equation (1) described above, where _Cave is (C [t ₁₀ ] + C [t ₉ ] + C [t ₈ ] + C [t ₇ ] + C [t ₆ ]) / 5. And the measured value were used for evaluation.

In the conventional single machine learning block, the target information is set to (C [t ₁₀ ] + C [t ₉ ] + C [t ₈ ] + C [t ₇ ] + C [t ₆ ]) / 5, and this time width is set. Trained to make predictions throughout. This corresponds to the throughput after 1.0 to 2.0 seconds, and in equation (1), the plot for 10 hours corresponds to approximately 180,000 points (n = 180000). Further, the measured value and the predicted value are set as C of the equation (1), respectively, with the throughput C [t ₆ ] after 1.0 to 1.2 seconds and the throughput C [t ₁₀ ] after 1.8 to 2.0 seconds, respectively. And the R2 score was calculated.

In FIG. 14, the results when the input information is {X [t], Y [t], kz [t], w [t], Vx [t], V _θ [t]} are compared. .. In predicting the throughput after 1.0 to 2.0 seconds, the result was 0.642948 for a single machine learning block. In addition, since the model of a single machine learning block can obtain only the output for a single time, if you want to obtain the throughput after 1.8 to 2.0 seconds, you have to generate a dedicated model separately. Must be.

On the other hand, looking at the results of using the first machine learning block 501 and the second machine learning block 502 in FIG. 14, the R2 score is 0.647799, which is approximately more accurate than that of a single machine learning block. It has improved by 0.7%.

It is also possible to use predicted values for individual times. For example, in the example of FIG. 14, the R2 score at a short time interval is shown for the throughput C [t ₁₀ ] after 1.8 to 2.0 seconds and the throughput C [t ₆ ] after 1.0 to 1.2 seconds. Has been done. For predictions at such short time intervals, the R2 score is low. In this way, it is possible to predict the throughput corresponding to an arbitrary time width such as 0.4 seconds or 0.8 seconds.

Further, in FIG. 14, looking at the results using the third machine learning block 601 and the fourth machine learning block 603, the R2 score is 0.672348, which is improved from the first configuration, which is a conventional simple. The accuracy is improved by 4.5% compared to one machine learning block. Further, the prediction performance for the corresponding throughput after 1.0 to 1.2 seconds and 1.8 to 2.0 seconds is also obtained from the first configuration by the first machine learning block 501 and the second machine learning block 502. Is also improving.

Further, the feature amount to be input is {X [t], Y [t], kz [t], w [t], Vx [t], V _θ [t], Ox [t], O _θ [t]}. In the case of 8 types, the prediction performance is higher than that of the conventional single machine learning block, but the prediction performance is improved compared to the case of 6 types of feature quantities to be input. There is only a second configuration using the machine learning block 601 and the fourth machine learning block 603. This is because the second configuration using the third machine learning block 601 and the fourth machine learning block 603 represents a more complicated latent structure than the other configurations, and the number of features to be input has increased. It is considered that the correspondence has been completed.

(Other realization forms)
Here, a computer may execute a process in which each block or a part of the blocks of the base station device 101 or the terminal device 102 in each of the above-described embodiments is used as a component. In that case, a program for realizing the processing performed by each block or a part of the blocks is recorded on a computer-readable recording medium, and the program recorded on this recording medium is read by the computer system so that the computer can read the program. You may let it run.

Note that the "computer" includes hardware such as an OS and peripheral devices. The "computer-readable recording medium" is a portable medium such as a flexible disk, a magneto-optical disk, a ROM, or a CD-ROM, or a storage device such as a hard disk built in or externally connected to a computer system.

Furthermore, the "computer-readable recording medium" includes programs acquired via a network such as the Internet or a communication line such as a telephone line, and is a program that is retained only for a short time, a program that is dynamically retained, and the like. Is also included.

Further, it may include a memory that holds a program for a certain period of time, such as a volatile memory inside a computer system that is a server or a client. Further, the program may be for realizing a part of the above-mentioned components, and may be realized by combining the above-mentioned components with a program already recorded in the computer system. good.

Further, the program may be realized by using hardware such as PLD (Programmable Logic Device) or FPGA (Field Programmable Gate Array).

As described above, the communication information prediction device, the communication information prediction method, and the communication information prediction program according to the present invention use input information including environmental information and set communication-related parameters as time-series data under a plurality of time conditions. When outputting, it is possible to efficiently generate a machine learning model and predict communication information. In particular, by using a common machine learning block, the learning cost and the model usage cost can be reduced.

Although the embodiments of the present invention have been described with reference to the drawings, the specific configuration is not limited to the above-described embodiments, and includes designs and the like within a range that does not deviate from the basic features of the present invention. Will be.

100 ... wireless communication system; 101 ... base station device; 102 ... terminal device; 201 ... NW section; 202 ... communication section; 203 ... communication environment model generation section; 204 ... -Environmental information generation unit; 205 ... Model utilization unit; 301 ... Communication unit; 302 ... Communication environment model generation unit; 303 ... Environmental information generation unit; 304 ... Model utilization unit; 401. Communication environment model; 501 ... first machine learning block; 502 ... second machine learning block; 503, 604 ... target information; 601 ... third machine learning block; 602 ...・ ・ Feature space information; 603 ・ ・ ・ Fourth machine learning block

Claims (7)

1. In a communication information prediction device that generates a communication environment model related to wireless communication of a moving terminal device and predicts communication information of the terminal device.
   An environment information generation unit that generates environment information related to the environment of at least one of the terminal device and the wireless communication destination device of the terminal device, and
   A communication unit that generates communication information related to wireless communication of the terminal device, and
   With the communication environment model generator that generates a communication environment model that outputs communication information corresponding to multiple time conditions by sharing machine learning blocks composed of the same coefficients and structures using the input information including the environment information. ,
   A communication information prediction device comprising a model utilization unit that predicts communication information of the terminal device using the generated communication environment model.
2. In the communication information prediction device according to claim 1,
   The environmental information includes the position, attitude, state, movement, control command of the terminal device and the device of the wireless communication destination of the terminal device, information of a camera and a sensor connected to the terminal device or the network, and behavior of the terminal or network. A communication information predictor characterized by containing at least one piece of information about a strategy.
3. In the communication information prediction device according to claim 1 or 2.
   The communication environment model is
   The first machine learning block that inputs time-series input information and performs machine learning,
   A communication information prediction device comprising a second machine learning block that outputs communication information corresponding to a plurality of times by using the output of the first machine learning block.
4. In the communication information prediction device according to claim 1 or 2.
   The communication environment model is
   A third machine learning block that outputs feature space information corresponding to multiple times for time-series input information,
   It is characterized by having a fourth machine learning block that outputs communication information at the same time as the feature space information by using the feature space information corresponding to a plurality of times output from the third machine learning block. Communication information prediction device.
5. The communication information prediction device according to claim 4.
   The third machine learning block is
   The first machine learning block that inputs time-series input information and performs machine learning,
   Using the output of the first machine learning block, the feature space information corresponding to each of a plurality of times is generated, and the second machine learning block is output to the fourth machine learning block. Communication information prediction device.
6. It is a communication information prediction method that generates a communication environment model related to wireless communication of a moving terminal device and predicts the communication information of the terminal device.
   Environmental information generation processing for generating environmental information related to the device environment of at least one of the terminal device and the wireless communication destination device of the terminal device, and
   Communication processing that generates communication information related to wireless communication of the terminal device, and
   A communication environment that uses input information including the environment information to share a machine learning block composed of the same coefficient and structure, and generates a communication environment model that outputs target information including communication information corresponding to a plurality of time conditions. Model generation process and
   A communication information prediction method, characterized in that a model utilization process for predicting communication information of the terminal device is performed using the generated communication environment model.
7. A communication information prediction program that generates a communication environment model related to wireless communication of a mobile terminal device and causes a computer to execute a process of predicting the communication information of the terminal device.
   Environmental information generation processing for generating environmental information related to the device environment of at least one of the terminal device and the wireless communication destination device of the terminal device, and
   Communication processing that generates communication information related to wireless communication of the terminal device, and
   A communication environment that uses input information including the environment information to share a machine learning block composed of the same coefficient and structure, and generates a communication environment model that outputs target information including communication information corresponding to a plurality of time conditions. Model generation process and
   A communication information prediction program characterized in that a computer executes a model utilization process for predicting communication information of the terminal device using the generated communication environment model.

Images