JP7794439B2 - Estimation device, wireless communication system including the same, and program for causing a computer to execute the same - Google Patents

Description

This invention relates to an estimation device, a wireless communication system including the same, and a program for causing a computer to execute the same.

Next-generation communications standards require the creation of a communications infrastructure that will enable access to all people, information, and things anywhere, both domestically and internationally, while enhancing the high speed, large capacity, low latency, and multiple connection capabilities that are characteristic of fifth-generation mobile communications systems (5G).

High-frequency radio waves cannot be expected to propagate beyond line of sight, resulting in weak spots. For this reason, there has been active research into technology that solves the problem of high-frequency radio waves by using an intelligent reflecting surface (IRS), which is made up of multiple passive reflecting elements whose reflection characteristics can be dynamically changed (Non-Patent Document 1). By appropriately changing the reflection characteristics of each element and creating a radio wave propagation path that bypasses obstacles, stable communication can be achieved even beyond line of sight.

Because IRS essentially acts as an electromagnetic wave reflector and does not contain high-frequency circuits, it also has the advantage of consuming significantly less power than conventional multi-hop communications used to expand coverage.

Q. Wu and R. Zhang, “Towards Smart and Reconfigurable Environment: Intelligent Reflecting Surface Aided Wireless Network,” IEEE Commun. Mag., vol. 58, no. 1, pp. 106-112, 2020.

However, in wireless communication systems using IRS (IRS-assist systems), the achievable transmission rate is not predicted, nor is the number of idle slots, in which the IRS's reflecting elements are not in use, so it is difficult to determine in advance whether enough reflecting elements are available to ensure a sufficient transmission rate and to create a reflecting element usage schedule.

Accordingly, an embodiment of the present invention provides an estimation device that can create a schedule for using reflecting elements by determining in advance whether enough reflecting elements are actually available to ensure a sufficient transmission rate.

Furthermore, according to an embodiment of the present invention, a wireless communication system is provided that includes an estimation device that can determine in advance whether enough reflecting elements are actually available to ensure a sufficient transmission rate and create a reflecting element usage schedule.

Furthermore, according to an embodiment of the present invention, a program is provided that causes a computer to make a prior determination as to whether enough reflective elements are actually available to ensure a sufficient transmission rate and to create a schedule for using the reflective elements.

(Configuration 1)
According to an embodiment of the present invention, an estimation device is provided for estimating information useful for predicting a usage schedule that is a schedule for using N reflecting elements in a wireless communication system in which a transmitter having M _{T antennas} (M _T is an integer of 2 or more), a first IRS that is an electromagnetic wave reflector having N reflecting elements ₍ N is an integer of 2 or more) whose reflection characteristics can be dynamically changed, and a receiver having M R antennas (M R is an integer of 2 or more) are non-linearly arranged, the estimation device comprising: a collection means; and an estimation means. The collection means collects first time series data that indicates, for each of the N reflecting elements, time series data consisting of a busy state in which the reflecting element is in use and an idle state in which the reflecting element is not in use, and second time series data that indicates an achievable rate that is an achievable transmission rate of the N reflecting elements when the transmitter is transmitting a signal using the M _T antennas. The estimation means performs a first estimation process that uses a first learning device to estimate a usability prediction result for each of the N reflecting elements, which is the result of predicting whether the reflecting elements in the first IRS are usable or not, based on the first time series data, and a second estimation process that uses a second learning device to estimate the achievable rates of the N reflecting elements of the first IRS in the propagation path for the reflecting elements that are in use among the N reflecting elements, based on the second time series data.

(Configuration 2)
In configuration 1, the estimation means, in a first estimation process, inputs first time series data to a first learning device and estimates usable reflecting elements among the N reflecting elements by receiving from the first learning device an idle state corresponding to the maximum probability among probabilities that a first class classified according to the idle state of the first time series data matches a first learning stage class, which is a class classified according to the idle state of the first learning time series data acquired in the learning stage; and, in a second estimation process, inputs second time series data to a second learning device and estimates the achievable rate of a used reflecting element among the N reflecting elements by receiving from the second learning device an achievable rate corresponding to the maximum probability among probabilities that a second class classified according to the achievable rate of the second time series data matches a second learning stage class, which is a class classified according to the achievable rate of the second learning time series data acquired in the learning stage.

(Configuration 3)
In configuration 1 or 2, the first learning device consists of N first machine learning devices, the second learning device consists of E (E is the number of reflection patterns of one or more of the N reflection elements when one or more of the reflection elements are used simultaneously) second machine learning devices, the first time series data consists of first idle state/busy state data to Nth idle state/busy state data, each of which arranges idle states and busy states in time series, and the second time series data consists of first achievable rate data to Eth achievable rate data, each of which arranges achievable rates in time series, associated with the E reflection patterns.

The estimation means is
In a first estimation process, first idle state/busy state data is input to one first machine learning device, and a first process is performed to receive from the one first machine learning device an idle state corresponding to the maximum probability among the probabilities that a first class classified according to the idle state of the first idle state/busy state data matches a first learning stage class, which is a class classified according to the idle state of the first idle state/busy state data of the learning stage acquired in the learning stage, and N usability prediction results of N reflecting elements are estimated by performing the same process as the first process on the second idle state/busy state data to the Nth idle state/busy state data instead of the first idle state/busy state data in the first process and all of the (N-1) first machine learning devices;
In the second estimation process, the first achievable rate data is input into one second machine learning machine, and a second process is performed in which an achievable rate corresponding to the maximum probability among the probabilities that the first class classified according to the achievable rate of the first achievable rate data matches a second learning stage class, which is a class classified according to the achievable rate of the first achievable rate data of the learning stage obtained in the learning stage, is received from the one first machine learning machine, and the achievable rate of the N reflecting elements used is estimated by performing the same process as the second process for the second achievable rate data to the Eth achievable rate data and all (E-1) second machine learning machines, instead of the first achievable rate data in the second process.

(Configuration 4)
In configuration 3, N first machine learning machines execute a first process in parallel in a first estimation process, and E second machine learning machines execute a second process in parallel in a second estimation process.

(Configuration 5)
In configuration 3 or 4, each of the N first machine learning devices classifies the idle states acquired in the first learning stage into a first plurality of classes and stores them in a first prediction matrix, and when one piece of idle state/busy state data is input, selects one class from the first plurality of classes stored in the first prediction matrix that has the highest probability of matching the idle state of the input one piece of idle state/busy state data, and outputs a usability prediction result to the estimation means based on the selection of one class;
Each of the E second machine learning machines classifies the achievable rates obtained in the second learning stage into a second plurality of classes and stores them in a second prediction matrix. When a single achievable rate data is input, the machine learning machine selects one class from the second plurality of classes that has the highest probability of matching the achievable rate of the input single achievable rate data, and outputs the achievable rate classified into the selected class to the estimation means.

(Configuration 6)
In any of configurations 1 to 5, the estimation means, in the second estimation process, converts the propagation paths to the M _R antennas of the receiver into M _R independent propagation paths, calculates M _R transmission rates for the M _R independent propagation paths, and determines the sum of the calculated M _R transmission rates as the achievable rate.

(Configuration 7)
In configuration 6, the estimation means calculates M _R transmission rates based on the signal-to-interference-plus-noise power ratios of the M _R antennas in the second estimation process.

(Configuration 8)
According to an embodiment of the present invention, a wireless communication system includes a first estimator including the estimator according to any one of configurations 1 to 7, a transmitter, a receiver, and a first IRS. The first estimator is disposed within the first IRS.

(Configuration 9)
In configuration 8, the wireless communication system further includes a second estimator and a second IRS. The second estimator is the estimator according to any one of configurations 1 to 7. The second IRS is an electromagnetic wave reflector having N ^* (N ^* is an integer equal to or greater than 2) reflecting elements whose reflectivity can be dynamically changed. The second estimator is disposed within the second IRS.

(Configuration 10)
Further, according to an embodiment of the present invention, the program
A program for causing a computer to execute estimation of information useful for predicting a usage schedule, which is a schedule for using N reflecting elements, in a wireless communication system in which a transmitter having M _T (M _T is an integer of 2 or more) antennas, a first IRS which is an electromagnetic wave reflector having N (N is an integer of 2 or more) reflecting elements whose reflection characteristics can be dynamically changed, and a receiver having M _R (M _R is an integer of 2 or more) antennas are non-linearly arranged, the program comprising:
a first step in which a collecting means collects first time series data for each of the N reflecting elements, the first time series data being composed of a busy state in which the reflecting element is in use and an idle state in which the reflecting element is not in use, and second time series data for the N reflecting elements being an achievable rate, which is an achievable transmission rate of the N reflecting elements when a transmitter is transmitting a signal using the _M T antennas;
The program causes a computer to execute a second step in which an estimation means executes a first estimation process in which a usability prediction result is estimated for each of N reflecting elements, the usability prediction result being the result of predicting whether or not a reflecting element in a first IRS is usable using a first learning device based on one time series data, and a second estimation process in which a second learning device is used to estimate the achievable rates of the N reflecting elements of the first IRS in the propagation path for those reflecting elements that are in use among the N reflecting elements, based on second time series data.

(Configuration 11)
In configuration 10, the estimation means, in a first estimation process of the second step, inputs the first time series data to a first learning device and estimates usable reflecting elements among the N reflecting elements by receiving from the first learning device an idle state corresponding to the maximum probability among probabilities that a first class classified according to the idle state of the first time series data matches a first learning stage class, which is a class classified according to the idle state of the first learning time series data acquired in the learning stage; and, in a second estimation process of the second step, inputs the second time series data to a second learning device and estimates the achievable rate of a used reflecting element among the N reflecting elements by receiving from the second learning device an achievable rate corresponding to the maximum probability among probabilities that a second class classified according to the achievable rate of the second time series data matches a second learning stage class, which is a class classified according to the achievable rate of the second learning time series data acquired in the learning stage.

(Configuration 12)
In configuration 10 or 11, the first learning device consists of N first machine learning devices, the second learning device consists of E (E is the number of reflection patterns of one or more reflection elements when one or more of the N reflection elements are used simultaneously) second machine learning devices, the first time series data consists of first idle state/busy state data to Nth idle state/busy state data, each of which arranges idle states and busy states in time series, and the second time series data consists of first achievable rate data to Eth achievable rate data, each of which arranges achievable rates in time series, associated with E reflection patterns.

The estimation means is
In the first estimation process of the second step, the first idle state/busy state data is input to one first machine learning device, and a first process is performed to receive from the one first machine learning device an idle state corresponding to the maximum probability among the probabilities that a first class classified according to the idle state of the first idle state/busy state data matches a first learning stage class, which is a class classified according to the idle state of the first idle state/busy state data of the learning stage acquired in the learning stage, and the same process as the first process is performed for the second idle state/busy state data to the Nth idle state/busy state data instead of the first idle state/busy state data in the first process and all of the (N-1) first machine learning devices to estimate N usability prediction results of the N reflecting elements;
In the second estimation process of the second step, the first achievable rate data is input to one second machine learning machine, and a second process is performed to receive from the one first machine learning machine an achievable rate corresponding to the maximum probability among the probabilities that the first class classified according to the achievable rate of the first achievable rate data matches a second learning stage class, which is a class classified according to the achievable rate of the first achievable rate data of the learning stage obtained in the learning stage, and the achievable rate of the reflecting element being used among the N reflecting elements is estimated by performing the same process as the second process for the second achievable rate data to the Eth achievable rate data and all of the (E-1) second machine learning machines, instead of the first achievable rate data in the second process.

(Configuration 13)
In configuration 12, N first machine learning machines execute the first process in parallel in the first estimation process of the second step, and E second machine learning machines execute the second process in parallel in the second estimation process of the second step.

(Configuration 14)
In configuration 12 or 13, each of the N first machine learning devices classifies the idle states acquired in the first learning stage into a first plurality of classes and stores them in a first prediction matrix, and when one piece of idle state/busy state data is input, selects one class from the first plurality of classes stored in the first prediction matrix that has the highest probability of matching the idle state of the input one piece of idle state/busy state data, and outputs a usability prediction result to the estimation means based on the selection of one class;
Each of the E second machine learning machines classifies the achievable rates obtained in the second learning stage into a second plurality of classes and stores them in a second prediction matrix. When a single achievable rate data is input, the machine learning machine selects one class from the second plurality of classes that has the highest probability of matching the achievable rate of the input single achievable rate data, and outputs the achievable rate classified into the selected class to the estimation means.

(Configuration 15)
In any of configurations 10 to 14, the estimation means, in the second estimation process of the second step, converts the propagation paths to the M _R antennas of the receiver into M _R independent propagation paths, calculates M _R transmission rates for the M _R independent propagation paths, and sets the sum of the calculated M _R transmission rates as the achievable rate.

(Configuration 16)
In configuration 15, the estimation means calculates M _R transmission rates based on the signal-to-interference-plus-noise power ratios of the M _R antennas in the second estimation process of the second step.

You can create a schedule for using reflective elements by determining in advance whether enough reflective elements are actually available to ensure a sufficient transmission rate.

1 is a schematic diagram of a wireless communication system according to an embodiment of the present invention; FIG. 2 is a schematic diagram of the IRS controller shown in FIG. 1 . FIG. 10 is a conceptual diagram of time series data TSD_1(t). FIG. 10 is a conceptual diagram of time series data TSD_2(t). FIG. 3 is a schematic diagram of a machine learning device that constitutes the learning device 44 shown in FIG. 2 . FIG. 6 is a schematic diagram of the probabilistic neural network PNN ^{SLT —} 1 shown in FIG. 5. FIG. 10 is a conceptual diagram showing a prediction matrix P ^SLT _r that predicts usable reflective elements. FIG. 8 is a schematic diagram showing a specific example of the prediction matrix P ^SLT _r shown in FIG. 7 . FIG. 3 is a schematic diagram of a machine learning device that constitutes the learning device 45 shown in FIG. 2 . FIG. 10 is a conceptual diagram showing a prediction matrix P ^AR _r that predicts an achievable rate AR. 3 is a flowchart illustrating the operation of the estimation device shown in FIG. 2 . 12 is a flowchart for explaining the detailed operation of step S4 shown in FIG. 11. 12 is a flowchart for explaining the detailed operation of step S5 shown in FIG. 11. FIG. 1 is a schematic diagram of another wireless communication system according to an embodiment of the invention. 1 is a schematic diagram of learners 44 and 45 when an IRS controller 22 of an IRS 20 controls 2N reflecting elements 21. FIG. FIG. 15 is a schematic diagram illustrating one specific example of the wireless communication system shown in FIG. 14.

Embodiments of the present invention will be described in detail with reference to the drawings. Note that identical or corresponding parts in the drawings will be designated by the same reference numerals, and their description will not be repeated.

Figure 1 is a schematic diagram of a wireless communication system according to an embodiment of the present invention. Referring to Figure 1, the wireless communication system 10 according to the embodiment of the present invention includes a transmitter 1, an IRS 2, and a receiver 3. The IRS 2 includes N (N is an integer equal to or greater than 2) reflecting elements 21 and an IRS controller 22. The N reflecting elements 21 are arranged, for example, in a grid pattern. That is, the N reflecting elements 21 are arranged at equal intervals. The IRS controller 22 includes an estimation device 4.

Transmitter 1, IRS 2, and receiver 3 are arranged in a wireless communication space. Transmitter 1, IRS 2, and receiver 3 are arranged in a position where the arrangement [Transmitter 1 - IRS 2 - Receiver 3] is not on a straight line. More specifically, if the line connecting Transmitter 1 and IRS 2 is L1 and the line connecting IRS 2 and receiver 3 is L2, Transmitter 1, IRS 2, and receiver 3 are arranged so that the angle α formed by lines L1 and L2 is greater than 0. In other words, Transmitter 1, IRS 2, and receiver 3 are arranged non-linearly in the wireless communication space.

The transmitter 1 has M _T (M _T is an integer equal to or greater than 2) antennas, and the receiver 3 has M _R (M _R is an integer equal to or greater than 2) antennas. M _T may be the same as or different _from M _R.

The transmitter 1 transmits a signal using M _T antennas. In the IRS 2, the IRS controller 22 controls the reflection coefficients of the N reflecting elements 21 to control the reflection angle and phase of the radio waves by the N reflecting elements 21. The N reflecting elements 21 then reflect the radio waves incident from the transmitter 1 toward the receiver 3. As a result, the radio waves reflected by the N reflecting elements 21 reach the receiver 3.

Receiver 3 receives radio waves transmitted from transmitter 1. In wireless communication system 10, the propagation path between transmitter 1 and receiver 3 consists of a direct link, in which radio waves from transmitter 1 reach receiver 3 directly, and an indirect link, in which radio waves from transmitter 1 reach receiver 3 via IRS 2.

If there is no obstacle OBT between transmitter 1 and receiver 3, receiver 3 receives radio waves transmitted from transmitter 1 using a direct link and an indirect link.

On the other hand, if an obstacle OBT exists between transmitter 1 and receiver 3, receiver 3 receives the radio waves transmitted from transmitter 1 using an indirect link.

Therefore, in the wireless communication system 10, even if an obstacle OBT exists between the transmitter 1 and the receiver 3, radio waves can be transmitted from the transmitter 1 to the receiver 3 by passing through the IRS 2.

The receiver 3 receives the signal transmitted from the transmitter 1, detects the signal-to-noise ratio (SNR), which is the ratio of signal power to noise power when the signal is received, and transmits the detected signal-to-noise ratio (SNR) to the IRS controller 22.

The estimation device 4 collects, as time series data TSD_1(t), N pieces of time series data D1_chr_1(t) to D1_chr_N(t) showing time series data D1_chr(t) for N reflecting elements 21 of the IRS 2, which consists of a busy state B in which the reflecting elements 21 are in use and an idle state I in which the reflecting elements 21 are not in use.

Furthermore, the estimation device 4 collects, as time series data TSD_2( _t ), N pieces of time series data D2_chr_1(t) to D2_chr_N(t), which indicate, for N reflecting elements 21, the achievable transmission rate of one reflecting element when the transmitter 1 is transmitting a signal using M T antennas.

The estimation device 4 then estimates usability prediction results for the N reflecting elements 21 in the IRS 2, which are results of predicting whether each of the N reflecting elements 21 in the IRS 2 is usable or not, based on the N pieces of time series data D1_chr_1(t) to D1_chr_N(t) using a method described below.

In addition, the estimation device 4 estimates the achievable rate of one reflecting element 21 of the IRS2 for N reflecting elements 21 on the propagation path from the transmitter 1 to the receiver 3 via the IRS2, based on the N pieces of time series data D2_chr_1(t) to D2_chr_N(t), using a method described below.

Furthermore, the estimation device 4 estimates the frequency response of the propagation path between the transmitter 1 and the receiver 3 using a method described below. In this case, the transmitter 1 transmits a pilot signal, which is a known signal, and the estimation device 4 receives the received signal y(t) from the receiver 3 via the IRS controller 22 when the receiver 3 receives the pilot signal. The estimation device 4 then estimates the frequency response of the propagation path based on the received signal y(t).

Figure 2 is a schematic diagram of the IRS controller 22 shown in Figure 1. Referring to Figure 2, the IRS controller 22 includes an antenna 221, communication means 222, control means 223, and an estimation device 4. The estimation device 4 includes collection means 41, storage means 42, estimation means 43, and learners 44 and 45.

The communication means 222 of the IRS controller 22 receives the received pilot signal y(t) from the receiver 3 via the antenna 221 and outputs the received pilot signal y(t) to the estimation means 43 of the estimation device 4.

In addition, the communication means 222 of the IRS controller 22 receives the signal-to-noise ratio SNR from the receiver 3 via the antenna 221 when the receiver 3 receives the received signal y(t) of the pilot signal, and outputs the received signal-to-noise ratio SNR to the collection means 41 of the estimation device 4.

Furthermore, when the communication means 222 of the IRS controller 22 receives an instruction signal INS_1 from the control means 223 instructing the transmitter 1 to transmit a training signal, it transmits the instruction signal INS_1 to the transmitter 1 via the antenna 221.

When a signal is transmitted from the transmitter 1 to the receiver 3 via the IRS 2, the control means 223 of the IRS controller 22 detects, in time series, the period ON_period during which the reflective element 21 is ON and the period OFF_period during which the reflective element 21 is OFF for each of the N reflective elements. The control means 223 then generates periods ON_period/period OFF_period(t) in which the periods ON_period and periods OFF_period are arranged in chronological order for the N reflective elements 21 to generate N periods ON_period/period OFF_period_1(t) through period ON_period/period OFF_period_N(t), and outputs the generated N periods ON_period/period OFF_period_1(t) through period ON_period/period OFF_period_N(t) to the collection means 41 of the estimation device 4.

Furthermore, the control means 223 of the IRS controller 22 receives from the estimation means 43 of the estimation device 4 the maximum achievable rate R max among the achievable rates calculated by applying each reflected beamforming vector in the codebook to each of the combinations of available reflecting elements (combinations possible among reflecting elements predicted to be in the OFF state) among the combinations of the N reflecting elements 21 of the IRS 2, the combination of reflecting elements that gives the maximum achievable rate R _max , and the reflected beamforming vector φ _max . Then, the control means 223 of the IRS controller 22 tally up the maximum achievable rate R _max , the combination of _reflecting elements that gives the maximum achievable rate R _max , and the reflected beamforming vector φ _max .

The control means 223 of the IRS controller 22 aggregates the maximum achievable rate R _max , the combination of reflecting elements that gives the maximum achievable rate R _max , and the reflected beamforming vector φ _max , and holds [maximum achievable rate R _max / combination of reflecting elements that gives the maximum achievable rate R _max / reflected beamforming vector φ _max ] that mutually corresponds the maximum achievable rate R _max , the combination of reflecting elements that gives the maximum achievable rate R _max , and the reflected beamforming vector φ _max .

Then, when transmitting a signal from the IRS 2 to the receiver 3 at the maximum achievable rate R _max , the control means 223 of the IRS controller 22 controls the reflection angle and phase of the radio waves by the N reflecting elements 21 by controlling the reflection coefficients of the N reflecting elements 21 based on the reflection beamforming vector φ _max .

When the collection means 41 receives N periods ON_period/period OFF_period_1(t) through ON_period/period OFF_period_N(t) from the control means 223 of the IRS controller 22, it generates time series data for the N periods ON_period/period OFF_period_1(t) through ON_period/period OFF_period_N(t) in which the periods ON_period are in busy state B and the periods OFF_period are in idle state I, thereby generating time series data TSD_1(t) and storing the generated time series data TSD_1(t) in the storage means 42.

The collection means 41 also receives the signal-to-noise ratio SNR from the communication means 222 of the IRS controller 22, and receives the frequency response FQR_0 of the propagation path between the transmitter 1 and the receiver 3 from the estimation means 43. Each time the collection means 41 receives the signal-to-noise ratio SNR, it calculates the achievable rate AR based on the signal-to-noise ratio SNR and the frequency response FQR_0 using a method described below, generates time series data TSD_2(t) that associates the achievable rate AR with the calculation time t of the achievable rate AR, and stores the generated time series data TSD_2(t) in the storage means 42.

The storage means 42 stores the time series data TSD_1(t) and the time series data TSD_2(t).

The estimation means 43 receives the received signal y(t) from the communication means 222 of the IRS controller 22, estimates the frequency response FQR_0 based on the received signal y(t) using a method described below, and stores the estimated frequency response FQR_0 while outputting the frequency response FQR_0 to the collection means 41.

Furthermore, the estimation means 43 reads out the time series data TSD_1(t) from the storage means 42, inputs the read out time series data TSD_1(t) to the learning device 44, and receives N usability prediction results U ₁ (t) to U _N (t) from the learning device 44. Each of the usability prediction results U ₁ (t) to U _N (t) is a result of predicting whether one reflecting element 21 is usable or not.

Furthermore, the estimation means 43 reads out the time series data TSD_2(t) from the storage means 42, inputs the read out time series data TSD_2(t) to the learning device 45, and receives N achievable rates AR ₁ to AR _N from the learning device 45.

The learning device 44 receives the time series data TSD_1(t) from the estimation means 43, predicts N usability prediction results U ₁ (t) to U N (t) based on the received time series data _{TSD_1} (t) using a method described later, and outputs the predicted N usability prediction results U ₁ (t) to U _N (t) to the estimation means 43.

The learning device 45 receives the time series data TSD_2(t) from the estimation means 43, predicts N achievable rates AR ₁ to AR _N based on the received time series data TSD_2(t) using a method described below, and outputs the predicted N achievable rates AR ₁ to AR _N to the estimation means 43.

Frequency response estimation
Assuming that there is no variation in the propagation path while observing the received signal y(t) (assuming that the frequency response H does not have (t)), and since beamforming is not performed in the transmitter 1, if the transmission weight W in the transmitter 1 is omitted, the received signal y(t) can be expressed by the following equation.

In equation (1), Θ is a matrix representing the N reflection coefficients of the N reflecting elements 21, and n(t) represents white noise.

When M _T received signals y(t) are observed and the M _T observed received signals y(t) are arranged, the following equation is obtained.

Then, in equation (2), if we multiply Y(t) by the inverse matrix of X(t) from the right, we obtain the following equation:

In the last line of equation (3), H consists of [H ^D H ^C ], which includes the frequency response H ^D in the direct link and the frequency response H ^C in the indirect link, and Ω consists of a matrix in which an identity matrix I and Θ (the reflection coefficient matrix of N reflecting elements 21) are arranged vertically.

Then, when the modes of the N reflecting elements 21 are switched (i.e., when the patterns of the N reflection coefficients of the N reflecting elements 21 are switched), the following equation is obtained:

However, in equation (4), the relationship t = _iMT is satisfied. As a result, t represents the timing at which x(1) to x( _MT ) are transmitted in each of the periods PD_1 to _{PD_MR} shown in Fig. 4, and "i" in t = _iMT represents Ω(1) to Ω(N) shown in Fig. 4.

By observing N G(i)s in equation (4) and arranging the N G(i)s, the following equation is obtained.

Then, if we multiply g in equation (5) by the inverse matrix of R on the right, we get the following equation:

According to equation (5), R consists of [Ω(1)...Ω(N)], which is known because it consists of N modes of reflection coefficients (i.e., patterns of N reflection coefficients) of N reflecting elements 21. Therefore, the inverse matrix of R can be calculated.

Furthermore, g in equation (6) can be calculated using G(t) = Y(t)X ^-1 (t) in equation (3), where Y(t) is known because it is the signal received when the receiver 3 receives the pilot signal, and X(t) is known because it is the signal transmitted by the transmitter 1. As a result, "gR ^-1 " in equation (6) can be calculated.

Furthermore, "N" in equation (6) can be ignored when the amount of propagation path fluctuation is sufficiently greater than the noise power (the signal-to-noise ratio is high). When the amount of propagation path fluctuation is not sufficiently greater than the noise power, the noise level in propagation path estimation can be kept below a predetermined value by transmitting the same pilot signal multiple times in succession and averaging the received signals.

Therefore, by calculating gR ⁻¹ and NR ⁻¹ , the frequency responses H ^D and H ^C can be independently determined using equation (6). The frequency response H ^D is the frequency response of the direct link between the transmitter 1 and the receiver 3, and the frequency response H ^C is the frequency response of the indirect link between the transmitter 1 and the receiver 3.

As a result, using the above equations (1) to (6), the frequency response [H ^D H ^C ] of the propagation path between the transmitter 1 and the receiver 3 can be independently estimated based on the received signal y(t).

When the estimation means 43 of the estimation device 4 receives the received signal y(t), it estimates the frequency response [H ^D H ^C ] (=FQR_0) of the propagation path using equations (1) to (6) based on the received signal y(t).

[Collection of time-series data]
3 is a conceptual diagram of the time series data TSD_1(t), in which "I" represents an idle state and "B" represents a busy state.

Referring to Figure 3, each of the idle state I and busy state B is collected in slot (SLT) units.

The collecting means 41 collects time series data TSD_1(t) for the following (A) or (B).
(A) Collection of time series data TSD_1(t) when one or more of the N reflecting elements 21 are turned OFF simultaneously. (B) Classifying the N reflecting elements 21 into clusters of a certain number (for example, 9), and collecting the time series data TSD_1(t) by controlling the ON/OFF state of the reflecting elements on a cluster-by-cluster basis. In (B), the collection means 41 selects, for all clusters, the reflecting element located at the center of the arrangement positions of a certain number (two or more) of reflecting elements classified into one cluster as a representative reflecting element, and collects time series data indicating the ON/OFF state of the selected representative reflecting element to collect the time series data TSD_1(t).

Regarding the selection of the representative reflective element, the representative reflective element may be selected by a method other than the method of selecting the reflective element located at the center of the arrangement positions of a certain number of reflective elements classified into one cluster as the representative reflective element. For example, any reflective element from among a certain number of reflective elements classified into one cluster may be selected as the representative reflective element. Any method for selecting a representative reflective element may be used.

When collecting time series data TSD_1(t) in (A), the collection means 41 receives from the control means 223 of the IRS controller 22 a period ON_period/period OFF_period_n(t) in which a period OFF_period is a period during which one or more of the N reflecting elements 21 are turned OFF, and a period ON_period is a period during which all of the N reflecting elements 21 other than one or more reflecting elements 21 are turned ON, arranged in time series.

Then, the collection means 41 associates "I" with slot SLT_1 based on the period OFF_period, and "I" with slot SLT_2 based on the period ON_period, and continues in the same manner to generate the time series data TSD_1(t) shown in Figure 3.

In this case, the collection means 41 detects that the OFF reflective elements (one or more reflective elements 21) in slot SLT_1 are in an idle state I, and that the ON reflective elements (reflective elements other than one or more reflective elements 21) in slot SLT_2 are in a busy state B. Similarly, the collection means 41 detects that the OFF reflective elements (one or more reflective elements 21) in slots SLT_3, SLT_5, SLT_6, and SLT_9 are in an idle state I, and that the ON reflective elements (reflective elements other than one or more reflective elements 21) in slots SLT_4, SLT_7, and SLT_8 are in a busy state B, thereby generating time series data TSD_1(t).

As a result, the time series of idle state I and busy state B can be detected for each of the N reflective elements 21, and the time series data TSD_1(t) consists of N B/I ₁ (t) to B/I _N (t) that indicate the time series of idle state I and busy state B for each of the N reflective elements 21.

Then, the collection means 41 stores the generated time series data TSD_1(t) in the storage means 42.

Furthermore, when collecting time series data TSD_1(t) in (B), the collection means 41 detects that the cluster including the representative reflective element that is turned off in slot SLT_1 is in idle state I, and that the cluster including the representative reflective element that is turned on in slot SLT_2 is in busy state B. Similarly, it detects the idle state I for the cluster including the representative reflective element that is turned off in slots SLT_3, SLT_5, SLT_6, and SLT_9, and detects the busy state B for the cluster including the representative reflective element that is turned on in slots SLT_4, SLT_7, and SLT_8, thereby generating time series data TSD_1(t).

Then, the collection means 41 stores the generated time series data TSD_1(t) in the storage means 42.

Note that as long as a signal is being transmitted from the transmitter 1 to the receiver 3 via the IRS 2, the collection means 41 receives the period ON_period/period OFF_period_n(t) from the control means 223 of the IRS controller 22, and therefore continues to associate "I" or "B" with each slot SLT using the method described above.

As a result, the time series data TSD_1(t) represents the history of idle state I and busy state B (idle state I/busy state B).

Figure 4 is a conceptual diagram of the time series data TSD_2(t). The collection means 41 receives the frequency response matrix H(t) from the estimation means 43 and calculates the achievable rate R using the following method.

The N reflection coefficients θ and transmission weights W of the N reflecting elements are optimized using the following equation:

In equation (7), γ _mR is the signal to interference and noise ratio (SINR) of the M _R antennas, and is a function of the reflection coefficient θ and the transmission weight W. Also, ω _mR is a parameter representing the priority of each receiving antenna. In particular, if no priority is assigned to the antennas, ω ₁ = ω ₂ = ... = ω _MR = 1/M _R may be used.

Equation (7) represents the determination of the reflection coefficient θ and transmission weight W when the right-hand side is maximized.

Then, the M _R transmission rates for the M _R antennas are calculated as log ₂ (1+γ _mR ) in equation (7).

The transmitter 1 multiplies (precodes) the transmission signal of each antenna by a coefficient matrix based on the information on the propagation path matrix (= frequency response matrix H(t)) obtained from the estimation device 4, so that the propagation paths of the antenna pairs are independent.

Therefore, transmitter 1 determines the coefficient matrix using the following equation:

In equation (8), matrix T is a transmission weight matrix, and corresponds to the above-mentioned transmission weight W. Furthermore, matrix A is a K×K diagonal matrix, where K is the number of streams, and p ₁ , ..., p _K are transmission powers allocated to streams 1 to K, respectively.

When the transmitted signal s is multiplied by matrices T and A, the received signal y is expressed by the following equation:

Multiplying the received signal y by T ^H H ^H gives the following equation:

ΛA in the last line of equation (10) is a diagonal matrix. The results of dividing each diagonal element (M _R elements) of ΛA by the noise power (or estimated value of noise power) of each antenna are γ ₁ , γ ₂ , ..., γ _MR in equation (7).

Therefore, the collecting means 41 divides each diagonal element (M _R ) of ΨA by the noise power (or an estimated value of the noise power) of each antenna, and then substitutes the result of the division into γ _mR in equation (7) to calculate log ₂ (1 + γ _mR ) for all γ ₁ , γ ₂ , ..., γ _MR to calculate M _R transmission rates for the M _R antennas.

The collecting means 41 then determines the sum of the M _R transmission rates as the achievable rate R.

Multiplying (precoding) the transmission signal of each antenna by the coefficient matrix G shown in equation (8) makes the propagation paths of the antenna pair independent, so calculating the M _R transmission rates using equations (7) to (10) is equivalent to converting the propagation paths to the M _R antennas of the receiver 3 into M _R independent propagation paths and calculating the M _R transmission rates for the converted M _R independent propagation paths.

The collection means 41 collects the time series data TSD_2(t) shown in Figure 4 by plotting the achievable rate R against the time t at which the achievable rate R was calculated.

The time series data TSD_2(t) is data obtained by collecting the achievable rate R when one or more reflecting elements 21 that are turned on are used when the transmitter 1 is transmitting a signal using M _T antennas (when N reflecting elements 21 are clustered, the on/off patterns of all the reflecting elements 21 in the cluster are assumed to be the same), so the collection means 41 generates the time series data TSD_2_n(t) for all patterns of one or more reflecting elements 21 that are turned on. As a result, the time series data TSD_2(t) is made up of E pieces of time series data TSD_2_1(t) to TSD_2_E(t), where E is the number of patterns of one or more reflecting elements 21 that are turned on.

E is the number of reflection patterns of one or more reflective elements 21 when one or more of the N reflective elements 21 are used (ON), and is an integer equal to or greater than 2. For example, when reflective element 21_1 of the N reflective elements 21 is used, the reflection pattern of one or more reflective elements 21 is [reflective element 21_1 = ON, reflective element 21_2 = OFF, reflective element 21_3 = OFF, ..., reflective element 21_N = OFF]. When reflective elements 21_1 and 21_2 of the N reflective elements 21 are used, the reflection pattern is [reflective element 21_1 = ON, reflective element 21_2 = ON, reflective element 21_3 = OFF, ..., reflective element 21_N = OFF]. The same applies to the pattern of one or more reflective elements 21 when one or more reflective elements 21 other than reflective elements 21_1 and 21_2 of the N reflective elements 21 are simultaneously used.

Note that as long as a signal is being transmitted from the transmitter 1 to the receiver 3 via the IRS 2, the collection means 41 continues to calculate the achievable rate R using the method described above and plot the calculated achievable rate R against the calculation time t of the achievable rate R.

As a result, the time series data TSD_2(t) represents the history of the achievable rate R.

[Prediction of usability prediction results]
Fig. 5 is a schematic diagram of a machine learning device that constitutes the learner 44 shown in Fig. 2. Referring to Fig. 5, the learner 44 is made up of N probabilistic neural networks PNN ^SLT — 1 to PNN ^SLT — N.

As described above, the time series data TSD_1(t) is made up of N B/I ₁ (t) to B/I _N (t), and therefore the learning device 44 is made up of N probabilistic neural networks PNN ^SLT — 1 to PNN ^SLT —N.

The N probabilistic neural networks PNN ^SLT — 1 to PNN ^SLT — N receive the time series data TSD — 1 — 1(t) to TSD — 1_N(t) (= B/I ₁ (t) to B/I _N (t)), respectively.

Here, each of the time series data TSD_1_1(t) to TSD_1_N(t) consists of data in which the idle state I period OFF_period and the busy state B period ON_period are arranged in time series for one of the N reflecting elements 21.

Then, the N probabilistic neural networks PNN ^SLT _1 to PNN ^SLT _N predict whether or not each of the N reflecting elements 21 will be available for use in the upcoming slot SLT based on the time series data TSD_1_1(t) to TSD_1_N(t) (=B/I ₁ (t) to B/I _N (t)), respectively, and output the predicted availability prediction results U ₁ (t) to U _N (t) to the estimation means 43. In this case, each of the availability prediction results U ₁ (t) to U _N (t) consists of "0" indicating that one reflecting element 21 is in the idle state I, or "1" indicating that one reflecting element 21 is not in the idle state I.

Fig. 6 is a schematic diagram of the probabilistic neural network PNN ^SLT_1 shown in Fig. 5. Referring to Fig. 6, the probabilistic neural network PNN ^SLT_1 comprises an input layer, a hidden layer, a summation layer, and a decision layer.

The input layer consists of Q neurons (Q is an integer greater than or equal to 2). The hidden layer consists of M groups (M is an integer greater than or equal to 2), each of which consists of Q neurons. The sum layer consists of M neurons.

Each of the Q neurons in the input layer is connected to each of the QxM neurons in the hidden layer.

In the hidden layer, each group of Q neurons x _1,1 to x _1,Q is connected to one neuron in the sum layer.

M neurons in the summation layer are connected to one neuron in the decision layer.

The input layer normalizes the input data and outputs the normalized data to the hidden layer. In this case, each of the Q neurons in the input layer outputs normalized data to each of the Q x M neurons in the hidden layer.

The hidden layer calculates the Euclidean distance between the time series data TSD_1_COL(t) collected by the collection means 41 and the time series data TSD_1_SAV(t) (time series data TSD_1(t) collected in the learning stage) stored in the hidden layer. The hidden layer then outputs the calculated Euclidean distance to the summation layer. In this case, in the hidden layer, Q neurons x _1,1 to x _1,Q of the class 1 group output the Euclidean distance to one neuron in the summation layer associated with the class 1 group. Similarly, Q neurons x _1,1 to x _1,Q of the class M group output the Euclidean distance to one neuron in the summation layer associated with the class M group.

The summation layer sums the contributions of each class c to produce a probability given by:

In equation (11), x is the input vector, _xc,j is the jth training vector, _NR is the total number of training vectors, and _σc is the standard deviation of class c. Furthermore, _fc represents the probability that the collected vector x matches the class _c1learn of the data (time series data TSD_1(t) in the learning stage) stored in the prediction matrix P ^SLTr (the prediction matrix held in the summation _layer ⁾ .

The summation layer then outputs the probability generated by equation (11) to the decision layer. In this case, each of the M neurons in the summation layer outputs a probability generated by equation (11) to one neuron in the decision layer.

The decision layer detects the maximum probability among the M probabilities received from the M neurons in the summation layer, and selects and outputs class c _{COL_1} when the maximum probability is obtained. Class c _{COL_1} is a class classified according to the number of idle states I in the time series data TSD_1(t), and includes the number of idle states I in the time series data TSD_1(t). Class c _{COL_1} constitutes the "first class."

The data stored in the prediction matrix P ^SLT _r (time series data TSD — 1(t) in the learning stage) is the time series data TSD — 1_learn(t) acquired in the learning stage of the probabilistic neural network PNN — 1.

During the learning phase, the control means 223 of the IRS controller 22 instructs the transmitter 1 to periodically (e.g., once every 1 to 10 minutes) transmit a training signal to the IRS 2.

The transmitter 1 transmits the training signal to the IRS 2 in response to an instruction from the control means 223. In this case, the transmitter 1 transmits the training signal to the IRS 2 using M _T antennas.

Then, in the IRS 2, the control means 223 controls the reflected beamforming vectors φ of the N reflecting elements 21 so as to turn on one or more of the N reflecting elements 21 that are in the OFF state, and the N reflecting elements 21 control the reflection angle and phase of the radio waves in accordance with the control from the control means 223 to reflect the training signal in the direction of the receiver 3. In this case, the control means 223 holds a codebook, and selects V reflected beamforming vectors φ ₁ to _φV from the codebook for controlling the reflection patterns of the N reflecting elements 21 to V (V is an integer of ₂ or more) reflection patterns, and switches the reflected beamforming vectors of the N reflecting elements 21 to one reflected beamforming vector φ _n (n is any of 1 to V) of the selected _V reflected beamforming vectors φ 1 to φV for all of the V reflected beamforming vectors φ ₁ to _φV .

Each of the V reflected beamforming vectors φ ₁ to φ _V is a reflected beamforming vector in which one or more of the N reflecting elements 21 are simultaneously turned ON. Furthermore, selection of one reflected beamforming vector φ _n from the V reflected beamforming vectors φ ₁ to φ _V is performed in accordance with a predetermined order.

Therefore, the control means 223 selects V reflected beamforming vectors φ ₁ to φ _V from the codebook.

Then, the control means 223 of IRS2 acquires data D_ON/OFF_t consisting of a time series of _ON / _OFF for the N reflecting elements 21 when the reflection angle and phase of the radio waves by the N reflecting elements 21 are switched using V reflection beamforming vectors φ 1 to φ V, and acquires time series data D_I/B_1(t) to D_I/B_N(t) consisting of idle state I/busy state B for the N reflecting elements 21 based on the acquired data D_ON/OFF_t.

The control means 223 of the IRS 2 then outputs the acquired time series data D_I/B_1(t) to D_I/B_N(t) to the estimation means 43 of the estimation device 4 as time series data TSD_1_learn(t).

The estimation means 43 of the estimation device 4 receives the time series data TSD_1_learn(t) (= time series data D_I/B_1(t) to D_I/B_N(t)) from the control means 223, detects usable reflecting elements among the N reflecting elements 21 based on the received time series data TSD_1_learn(t) (= time series data D_I/B_1(t) to D_I/B_N(t)), classifies the usable reflecting elements into n _R (n _R is an integer of 2 or greater) classes c ¹ _{learn_1} to c ¹ _{learn_nR} according to the detected number of usable reflecting elements, and stores the n _R classes c ¹ _{learn_1} to c ¹ _{learn_nR} in n _R rows of the prediction matrix P ^SLT _r in the PNN ^SLT _1, respectively. Each of the _nR classes c ¹ _{learn —} 1 to c ¹ _{learn_nR} constitutes a “first learning stage class”.

In the learning stage, the above operations are performed periodically, so that the _nR classes ^c1learn_1 to ^c1learn_nR stored in the _nR rows of the prediction matrix P ^SLT _{r in} the PNN ^SLT_1 _are updated as needed.

7 is a conceptual diagram showing a prediction matrix P ^SLT _r that predicts available reflecting elements. Referring to FIG. 7, the time series data TSD_1(t) includes "■" and "□". "■" represents a busy state B, and "□" represents an idle state I. Each of the "■" and "□" is located in one slot SLT.

The prediction matrix P ^SLT _r consists of a matrix with _nR rows and _nC columns. The sliding window SDW1 has a window span n _C . ts(s).

The sliding window vector s _n , which is the vector of the sliding window SDW1, forms one row in the prediction _matrix ^PSLTr . As the sliding window SDW1 moves to the right with a window span _nC.ts (s) (see the arrow in Figure 7), one sample (one sample of the time series data TSD_1(t)) consisting of one new idle state I/busy state B is dropped into one row of the prediction _matrix ^PSLTr . As the sliding window SDW1 moves to the right, the sliding window vector s _n is used to drop one sample of the time series data TSD_1(t) into one row of the prediction matrix ^PSLTr , _{and nR samples} of the time series data TSD_1(t) are dropped into the _nR rows of the prediction _matrix ^PSLTr .

Each row T of the _nR x 1 target matrix contains a positive integer indicating an achievable usable reflective element class c (1 < T < K), where K is the total number of achievable usable reflective element classes c and is an integer greater than or equal to 2.

The prediction matrix P ^SLT _r performs the process of outputting, for nR rows, the probability that a class c _{COL —} 1 classified according to the idle state I of samples dropped in one row using the sliding window SDW1 (samples dropped by sliding window vectors s _n to s _n-nC+1, etc.) matches a class c ¹ _learn classified according to the idle state I of time series data stored in one row (time series data TSD — 1_learn( _t ) obtained in the learning stage).

Then, the target matrix outputs a positive integer (a positive integer indicating an idle state I or a busy state B) contained in row T that has received the maximum probability among the _nR probabilities received from the _nR rows of the prediction _matrix ^PSLTr .

In this way, the probabilistic neural network PNN ^SLT — 1 predicts, for one reflective element, the idle state I or the busy state B that is closest to the pattern history (time series data TSD — 1 (t)) using the prediction ^matrix P SLT _r .

As a result, when a piece of time series data TSD_1(t) is input, the probabilistic neural network PNN ^SLT _{_1} shown in FIG. 5 outputs the class c max (i.e., the availability prediction result U 1 ( ^t ) consisting of an idle state I or a busy state B) when the probability that the input piece of time series data TSD_1(t) matches the class c _learn of the data (time series data _{TSD_1_learn} (t)) stored in the prediction matrix P _SLT r (the prediction matrix held in the summation layer) is the highest.

Each of the probabilistic neural networks PNN ^SLT _2 to PNN ^SLT _N shown in Fig. 5 has the same configuration as the probabilistic neural network PNN ^SLT _1 shown in Fig. 6. Therefore, when time-series data B/I ₂ (t) to B/I _N (t) are input, the probabilistic neural networks PNN ^SLT _2 to PNN ^SLT _N output usability prediction results U ₂ (t) to U _N (t) for the N-1 reflecting elements, respectively.

This makes it possible to predict which of the N reflecting elements 21 can be used.

As a result, a communication schedule can be created using the predicted available reflecting elements.

Fig. 8 is a schematic diagram showing a specific example of the prediction matrix P ^SLT _r shown in Fig. 7. In Fig. 8, "0" represents the idle state I, and "1" represents not being in the idle state I.

Referring to FIG. 8, using the sliding window vector s _n shown in FIG. 7, one sample of the time series data TSD_1(t) is dropped into one row of the prediction matrix P ^SLT _r , and 100 (=n _R ) samples of the time series data TSD_1(t) are dropped into 100 (=n _R ) rows of the prediction matrix P ^SLT _r .

Thereby, for one reflective element, the idle/busy patterns I/B_PTN older than the time instant t _PDC are dropped into 100 (=n _R ) rows of the prediction matrix P ^SLT _r .

At time t _PDC when it is predicted whether one reflecting element 21 is usable or not, it is predicted whether one reflecting element 21 is usable or not.

As a result, "0" is stored in the "prediction result of usability" for Sequence 1 to Sequence 3, and "1" is stored in the "prediction result of usability" for Sequence 1 to Sequence 3.

For sequences 1 to 3, the predicted availability result is "0", so one reflecting element is predicted to be in an idle state in slot SLT after time t _PDC .

On the other hand, for sequence 100, the prediction result of availability is "1", so it is predicted that one reflective element is not in the idle state in slot SLT after time t _PDC .

Each of the usability prediction results for Sequences 1 to 100 is expressed as a probability, so of the 100 usability prediction results for Sequences 1 to 100, the usability prediction result with the highest probability is ultimately selected.

Therefore, when the predicted result of the final selected availability is an idle state (0), one reflective element is predicted to be available, and when the predicted result of the final selected availability is not an idle state (0), one reflective element is predicted to be unavailable.

Therefore, the N probabilistic neural networks PNN ^SLT — 1 to PNN ^SLT — N shown in FIG. 5 each predict whether or not the N reflecting elements 21 are available for use in the slot SLT after the time t _PDC , and output the prediction results to the estimation means 43.

Prediction of Achievable Rate AR
Fig. 9 is a schematic diagram of a machine learning device that constitutes the learning device 45 shown in Fig. 2. Referring to Fig. 9, the learning device 45 is made up of E probabilistic neural networks PNN ^AR — 1 to PNN ^AR — E.

The E probabilistic neural networks PNN ^AR — 1 to PNN ^AR — E receive the time series data TSD — 2 — 1 (AR ₁ (t) — _COL ) to TSD — 2 — E (AR _E (t) — _COL ) from the estimation means 43, respectively.

Then, the E probabilistic neural networks PNN ^AR — 1 to PNN ^AR — E estimate achievable rates AR ₁ (t) to ARE (t) based on the time series data TSD — 2 — 1 (AR ₁ (t) — _COL ) to TSD — ₂ — E ( _ARE (t) — _COL ), respectively, and output the estimated achievable rates AR ₁ (t) to _ARE (t) to the estimation means 43.

Each of the E probabilistic neural networks PNN ^AR — 1 to PNN ^AR — E has the same configuration as the probabilistic neural network PNN ^SLT — 1 shown in FIG.

In each of the E probabilistic neural networks PNN ^AR — 1 to PNN ^AR _E, the input layer normalizes input data and outputs the normalized data to the hidden layer, where each of the Q neurons in the input layer outputs normalized data to each of the Q×M neurons in the hidden layer.

The hidden layer calculates the Euclidean distance between the collected time series data TSD_2_COL(t) and the time series data TSD_2_SAV(t) stored in the hidden layer. The hidden layer then outputs the calculated Euclidean distance to the summation layer. In this case, in the hidden layer, Q neurons _x1,1 to _x1,Q of the class 1 group output the Euclidean distance to one neuron associated with the class 1 group in the summation layer. Similarly, N neurons _x1,1 to _x1,Q of the class M group output the Euclidean distance to one neuron associated with the class M group in the summation layer.

The summation layer sums the contributions of each class c to generate the probability represented by equation (11). When a probabilistic neural network PNN is used to predict the achievable rate AR, f _c in equation (11) represents the probability that the collected vector x matches the class c ² _learn of the data (time series data TSD_2(t)) stored in the prediction matrix P ^AR _r (the prediction matrix held in the summation layer).

The summation layer then outputs the probability generated by equation (11) to the decision layer. In this case, each of the M neurons in the summation layer outputs a probability generated by equation (11) to one neuron in the decision layer.

The decision layer detects the maximum probability among the M probabilities received from the M neurons in the summation layer, and selects and outputs the class c ² _learn when the maximum probability is obtained.

The data (time series data TSD_2(t)) stored in the prediction matrix P ^AR _r is the time series data TSD_2_learn(t) acquired in the learning stage of each of the E probabilistic neural networks PNN ^AR — 1 to PNN ^AR _E.

During the learning phase, the control means 223 of the IRS controller 22 instructs the transmitter 1 to periodically (e.g., once every 1 to 10 minutes) transmit a training signal to the IRS 2.

The transmitter 1 transmits the training signal to the IRS 2 in response to an instruction from the control means 223. In this case, the transmitter 1 transmits the training signal to the IRS 2 using M _T antennas.

In the IRS 2, the control means 223 uses the reflected beamforming vector φ to control the reflection angle and phase of the radio waves by the N reflecting elements 21, and the N reflecting elements 21 change the reflection angle and phase of the radio waves in accordance with the control from the control means 223 to reflect the training signal in the direction of the receiver 3. In this case, the control means 223 selects V reflected beamforming vectors φ _{1 to φ V} from the codebook, and performs this switching of the reflected beamforming vectors of the N reflecting elements 21 to one reflected beamforming vector φ _m (m is any of 1 to V) of the selected V reflected beamforming vectors φ ₁ to φ _V for all of the V reflected beamforming vectors φ 1 to φ V. Each of the V reflected beamforming vectors φ ₁ to φ _V is a reflected beamforming vector for controlling the N reflecting elements 21 so that one or more of the N reflecting elements 21 are turned ON and the reflecting elements 21 other than the one or more reflecting elements 21 are turned OFF. Furthermore, the selection of one reflected beamforming vector φ _m from the V reflected beamforming vectors φ ₁ to φ _V is performed in accordance with a predetermined order.

Thereafter, when the reflected beamforming vectors of the N reflecting elements 21 are switched to one reflected beamforming vector _φm , the estimation means 43 divides each diagonal element (M _R ) of ΨA shown in equation (10) by the noise power (or estimated value of noise power) of each antenna, and substitutes the result of the division into _γmR in equation (7) to calculate _log2 (1+ _γmR ), thereby calculating M _{R transmission} rates for all _γ1 , _γ2 , ..., _γMR .

The estimation means 43 then calculates the sum of the M _R transmission rates as the achievable rate R.

The estimation means 43 calculates the achievable rate R for all V reflected beamforming vectors φ ₁ to φ _V to calculate E achievable rates AR ₁ to _ARE .

Then, the estimation means 43 classifies the achievable rates AR ₁ to AR _E into n _R classes c ² _{learn —} 1 to c ² _{learn_nR} according to the value of the achievable rate AR, and stores the classified n R classes c ² _{learn —} 1 to c ² _{learn_nR} in n _R rows of the prediction matrix P ^AR _r in PNN ^AR _m (m is any one of ₁ to E) for all PNN ^AR _1 to PNN ^AR _E.

During the learning phase, the above operations are performed periodically, so that the nR classes ^c2learn_1 to ^c2learn_nR stored in the _nR rows of _the prediction matrix ^PARr in each of the _{E probabilistic} neural networks ^PNNAR_1 to ^PNNAR_E are updated as _needed .

10 is a conceptual diagram showing a prediction matrix P ^AR _r for predicting the achievable rate AR. Referring to FIG. 10, the time series data TSD_ 2 (t) is time series data collected by the collecting means 41.

The prediction matrix P ^AR _r consists of a matrix with _nR rows and _nC columns. The sliding window SDW2 has a window span n _C . ts(s).

The sliding window vector s _n , which is the vector of the sliding window SDW2, forms one row in the prediction matrix P ^AR _r . As the sliding window SDW2 moves to the right with a window span n _C .ts(s) (see the arrow in Figure 10), one sample (one sample of the time series data TSD_2(t)) consisting of one new achievable rate AR is dropped into one row of the prediction matrix P ^AR _r . As the sliding window SDW2 moves to the right, the sliding window vector s _n is used to drop one sample of the time series data TSD_2(t) into one row of the prediction matrix P ^AR _r , and _nR samples of the time series data TSD_2(t) into the _nR rows of the prediction matrix P ^AR _r .

Each row T of the n _R ×1 target matrix contains a positive integer indicating a class c ² _learn of the achievable rate AR (1≦T≦K).

The prediction matrix P ^AR _r performs this process for nR rows, outputting the probability that class c _{COL_2} of samples dropped into one row using the sliding window SDW2 (samples dropped by sliding window vectors s _n to s n- _{nC +1,} etc.) matches class c ² _learn of the time series data stored in one row (time series data TSD_2_learn(t) acquired in the learning stage). Note that class c _{COL_2} constitutes the "second class," and class c ² _learn constitutes the "second learning stage class."

Then, the target matrix outputs a positive integer (a positive integer indicating the achievable rate AR) contained in the row T that has received the maximum probability among the _nR probabilities received from the _nR rows of the prediction matrix P ^AR _r .

In this way, each of the E probabilistic neural networks PNN ^AR — 1 to PNN ^AR — E uses the prediction matrix P ^AR _r to find the maximum achievable rate AR that is closest to the pattern history (time series data TSD — 2(t)).

As a result, when a piece of time series data TSD_2(t) is input, each of the E probabilistic neural networks ^{PNN AR} _1 to PNN ^AR _E outputs an integer (i.e., any of the achievable rates AR 1 (t) to AR _E (t)) indicating the class c 2 max when the maximum probability is obtained that the class c _{COL_2} of the input piece of time series data TSD_2(t) matches the class c ² _learn of _the data (time series data _{TSD_2_learn} (t)) stored in the prediction matrix _P AR r (the prediction matrix held in the summation ^layer ).

Therefore, the achievable rate AR when the transmitter 1 transmits a signal using the M _T antennas can be predicted for one or more of the N reflecting elements 21 that are used simultaneously.

The estimation means 43 receives E achievable rates AR ₁ (t) _{_PRD} to ARE (t) _{_PRD} from E probabilistic neural networks PNN ^AR — 1 to PNN ^AR — _E. Then, based on the E achievable rates AR ₁ (t) — _PRD to AR _e′ (t) — _PRD , the estimation means 43 detects the achievable rates AR ₁ (t) — PRD to AR _e′ (t) — _PRD of the reflecting elements 21 predicted to be available for use in the future (one or more reflecting elements 21 for which the predicted availability result is “0”), and detects the maximum achievable rate AR _max of the detected achievable rates AR ₁ (t) — _PRD to AR _e′ (t) — _PRD and _the reflected beamforming vector φ _max of the N reflecting elements 21 when the maximum achievable rate AR _max is obtained, where e′ is in the range of 1≦e′≦E.

As a result, a communication schedule can be created using the maximum achievable rate AR _max .

The maximum achievable rate AR _max is the maximum achievable rate among the E achievable rates AR ₁ to ARE E estimated by switching one reflected beamforming vector φ _m , in which one or more of the N reflecting elements 21 are simultaneously turned ON, for all of the V reflected beamforming vectors φ ₁ to φ _V. Therefore, each of the E achievable rates AR ₁ to _{ARE E is} the achievable rate for a reflecting element that is being used among the N reflecting elements 21.

[Estimation of Reflection Beamforming Vector]
The codebook is designed by using the discrete Fourier transform of the reflection beamforming vector φ of the reflection element, and has a size of, for example, 32, 64, ..., 1024. The size is a parameter.

Reflective elements other than the newly turned ON will be set to "0".

The codebook DFT- _MH , designed using the Discrete Fourier Transform (DFT), is an azimuthal dimension codebook, where "M" in " _MH " represents the horizontal direction.

The m _H- th column element is expressed by the following equation:

In equation (12), m _H =1, 2, . . . , M _H , and j represents the imaginary unit.

For the case where M _H =4, the codebook DFT-M _H is shown in Table 1.

Since one or more of the N reflecting elements 21 that are used simultaneously are one or more reflecting elements that achieve the maximum achievable rate R _max , the estimation means 43 estimates the reflected beamforming vector by selecting from the codebook the reflected beamforming vector that achieves the maximum achievable rate R _max (the reflected beamforming vector for the one or more reflecting elements 21 that are used simultaneously).

As described above, the estimation means 43 of the estimation device 4 receives the usability prediction results U ₁ (t) to U _N (t) for the N reflecting elements 21 from the learning device 44, and also receives the E achievable rates AR ₁ (t) to _ARE (t) from the learning device 44.

Then, based on the usability prediction results U ₁ (t) to U _N (t), the estimation means 43 detects usable reflecting elements among the N reflecting elements 21. For example, the estimation means 43 detects two reflecting elements 21_1 and 21_2 as usable reflecting elements among the N reflecting elements 21.

Then, when the estimation means 43 detects two reflecting elements 21_1 and 21_2, it obtains three combinations of the two reflecting elements 21_1 and 21_2 when they are used: "reflecting element 21_1 only," "reflecting element 21_2 only," and "reflecting element 21_1 and reflecting element 21_2."

Then, the estimation means 43 detects the achievable rate AR _{21_1} when using "only reflective element 21_1", the achievable rate AR _{21_2} when using "only reflective element 21_2", and the achievable rates AR 21_1, AR _{21_2} when using "reflective element 21_1 and reflective element 21_2" from the E achievable rates AR ₁ (t) to AR _E (t).

Since the E achievable rates AR ₁ (t) to AR _E (t) are the achievable rates for all combinations when using one or more of the N reflective elements 21, the estimation means 43 can detect the achievable rate AR _{21_1} , the achievable rate AR _{21_2} and the achievable rate AR _{21_1, 21_2} from the E achievable rates AR ₁ (t) to AR _E (t).

When the estimation means 43 detects the achievable rate AR _{21_1} , the achievable rate AR _{21_2} and the achievable rate AR _{21_1, 21_2 ,} it selects the maximum achievable rate AR _max among the achievable rate AR _{21_1} , the achievable rate AR _{21_2} and the achievable rate AR 21_1, 21_2, and selects from the codebook the reflected beamforming vector φ _max at which the maximum achievable rate AR _max is obtained.

Here, it is assumed that the estimation means 43 detects the achievable rates AR _{21_1, 21_2} as the maximum achievable rate AR _max .

The estimation means 43 then creates a communication schedule based on the usable reflecting elements 21_1, 21_2 out of the N reflecting elements 21, the maximum achievable rate AR _max (=AR _{21_1, 21_2} ), and the reflected beamforming vector φ _max .

This allows you to create a schedule for using reflective elements by determining in advance whether enough reflective elements are actually available to ensure a sufficient transmission rate.

Fig. 11 is a flowchart for explaining the operation of the estimation device 4 shown in Fig. 2. Referring to Fig. 10, when the operation of the estimation device 4 is started, the estimation means 43 of the estimation device 4 estimates the frequency response matrices H _C and H _D of the propagation path from the transmitter 1 to the receiver 3 using equations (1) to (6) based on the received signal y(t) received from the receiver 3 (step S1).

Then, the collection means 41 of the estimation device 4 receives from the control means 223 the ON/OFF periods of the N reflecting elements detected by the control means 223, and collects time series data TSD_1(t) indicating the time series of the ON/OFF periods of the N reflecting elements based on the received ON/OFF periods of the N reflecting elements (step S2). The collection means 41 then stores the time series data TSD_1(t) in the storage means 42.

In this case, the control means 223 detects the periods during which one reflective element is ON and the periods during which one reflective element is OFF, and then arranges the detected ON and OFF periods in a time series to generate time series data for one reflective element, performing this process for all N reflective elements to generate time series data TSD_1(t). The control means 223 then outputs the time series data TSD_1(t) to the collection means 41, and the collection means 41 collects the time series data TSD_1(t) by receiving the time series data TSD_1(t) from the control means 223. The collection means 41 also stores the time series data TSD_1(t) in the storage means 42. Furthermore, the time series data TSD_1(t) consists of time series data D_I/B_1(t) to D_I/B_N(t) for the reflective elements 21_1 to 21_N, respectively.

After step S2, the collection means 41 calculates the achievable rate R using equations (7) to (10) using the frequency response matrix H(t) and collects time series data TSD_2(t) (step S3). The collection means 41 then stores the time series data TSD_2(t) in the storage means 42.

Subsequently, the estimation means 43 reads the time series data TSD_1(t) from the storage means 42, and estimates usable reflecting elements among the N reflecting elements 21 using the learning device 44 (PNN ^SLT _1 to PNN ^SLT _N) based on the read time series data TSD_1(t) (step S4).

Then, the estimation means 43 uses the learners 45 (PNN ^AR — 1 to PNN ^AR — E) to estimate E achievable rates R ₁ to R _E when one or more of the N reflecting elements are used simultaneously based on the time series data TSD_2(t) (step S5).

Thereafter, the estimation means 43 obtains the reflected beamforming vector φ _max of the N reflecting elements when the maximum achievable rate R _max among the achievable rates R ₁ to R _E is obtained (step S6).

Then, the estimation means 43 creates a communication schedule using the IRS 2 using the available reflecting elements among the N reflecting elements 21, the maximum achievable rate R _max among the E achievable rates R ₁ to R _E , and the reflected beamforming vector φ _max (step S7). This completes the operation of the estimation device 4.

In the flowchart shown in Fig. 11, a reflective beamforming vector φ _max of N reflective elements that achieves the maximum achievable rate R _max is obtained (see step S6), in order to maximize the transmission rate in the downlink.

Therefore, the estimation means 43 creates a communication schedule using the reflected beamforming vector φ _max (see step S7).

Figure 12 is a flowchart illustrating the detailed operation of step S4 shown in Figure 11.

12, after step S3 in FIG. 11, the probabilistic neural network PNN ^SLT — e (e is any one of 1 to E) of the learning device 44 receives the time series data D_I/B_e(t) from the estimation means 43 (step S41).

Then, the probabilistic neural network PNN ^SLT_e drops one sample consisting of _nC pieces of data extracted from the time series data D_I/B_e(t) using the sliding window vector s _n onto one row of the prediction matrix P ^SLT _r while moving the sliding window, thereby dropping _nC samples onto each of the _nR rows of the prediction matrix P ^SLT _r (step S42).

Then, the prediction matrix P ^SLT _r of the probabilistic neural network PNN ^SLT _e outputs the probability that the class c _{COL —} 1 _of the ^sample in one row of the prediction matrix P ^SLT _r matches the class c ¹ _learn of the learning sample stored in one row of the prediction matrix P SLT r, for all n _R rows (step S43).

Then, the target matrix T of the probabilistic neural network PNN ^SLT _n outputs the positive integer (positive integer indicating the prediction result of usability) contained in the row of the target matrix T that has received the maximum probability out of the _nR probabilities _{from the} ^prediction matrix P SLT r to the estimation means 43 (step S44). After step S44, the operation of the estimation device 4 proceeds to step S5 in FIG. 11.

In step S44 of FIG. 12, if there are multiple maximum probabilities among the _nR probabilities, the target matrix T outputs to the estimation means 43 a positive integer (a positive integer indicating the prediction result of usability) contained in the row of the target matrix T that has received any one of the multiple maximum probabilities.

12 in parallel and output U ₁ (t) to U _N (t) consisting of prediction results of usability to the estimation means 43. As a result, usable reflecting elements out of the N reflecting elements 21 are estimated by the probabilistic neural networks ^{PNN SLT _1} to PNN ^SLT _N.

Figure 13 is a flowchart illustrating the detailed operation of step S5 shown in Figure 11.

13, after step S4 in FIG. 11, the probabilistic neural network PNN ^AR — e (e is any one of 1 to E) of the learning device 45 receives the time series data TSD — 2 — e(t) from the estimation means 43 (step S51).

Then, the probabilistic neural network PNN ^AR_e drops one sample consisting of _nC pieces of data extracted from the time series data TSD_2_e(t) using the sliding window vector s _n into one row of the prediction matrix P ^AR _r while moving the sliding window, thereby dropping _nC samples into each of the _nR rows of the prediction matrix P ^AR _r (step S52).

Then, the prediction matrix P ^AR _r of the probabilistic neural network PNN ^AR _e outputs the probability that the class c ^COL _{— 2} of the _sample in one row of the prediction matrix P ^AR _r matches the class c ² _learn of the learning sample stored in one row of the prediction matrix P AR r for all _nR rows (step S53).

Then, the target matrix T of the probabilistic neural network PNN ^AR — e outputs the positive integer (positive integer indicating the achievable rate ^AR ) contained in the row of the target matrix T that has received the maximum probability out of the _nR probabilities from _the prediction matrix P AR r to the estimation means 43 (step S54). After step S54, the operation of the estimation device 4 proceeds to step S6 in FIG. 11.

In step S54 of FIG. 13, if there are multiple maximum probabilities among the _nR probabilities, the target matrix T outputs to the estimation means 43 a positive integer (a positive integer indicating the achievable rate AR) contained in a row of the target matrix T that has received any one of the multiple maximum probabilities.

13 in parallel to estimate the achievable rates ^AR 1 (t) to ^ARE (t), respectively. As a result, the E achievable rates AR ₁ (t) to _ARE (t) of the N reflective elements 21 are estimated by the probabilistic neural networks PNN ^AR _{_1} to PNN ^AR _{_E} .

According to the flowchart shown in Figure 11 (including the flowcharts shown in Figures 12 and 13), the estimation means 43 estimates usable reflecting elements among the N reflecting elements 21 and E achievable rates AR ₁ to ARE _E , and obtains the reflected beamforming vector φ _max of the N reflecting elements 21 when the maximum achievable rate AR _max among the E achievable rates AR ₁ to ARE _E is obtained.

Then, the estimation means 43 creates a communication schedule using the IRS2 using the available reflecting elements among the N reflecting elements 21, the maximum achievable rate AR _max among the _E achievable rates AR ₁ to ARE, and the reflected beamforming vector φ _max of the N reflecting elements 21.

The estimation means 43 then outputs to the control means 223 the usable reflecting elements among the N reflecting elements 21, the maximum achievable rate AR _max among the E achievable rates R ₁ to R _E , the reflected beamforming vectors φ _max of the N reflecting elements 21, and the communication schedule.

When the control means 223 receives the usable reflecting elements among the N reflecting elements 21, the maximum achievable rate AR _max among the _E achievable rates AR ₁ to ARE, the reflected beamforming vector φ _max of the N reflecting elements 21, and the communication schedule from the estimation means 43, it transmits the usable reflecting elements among the N reflecting elements 21, the maximum achievable rate AR _max among the _E achievable rates AR ₁ to ARE, and the communication schedule to the transmitter 1 and retains the reflected beamforming vector φ _max of the N reflecting elements 21.

The transmitter 1 receives from the control means 223 of the IRS 2 the available reflecting elements out of the N reflecting elements 21, the maximum achievable rate AR _max out of the _E achievable rates AR ₁ to ARE, and the communication schedule.

Then, the transmitter 1 refers to the N reflecting elements 21 that are available and the maximum achievable rate AR _max among the _E achievable rates AR ₁ to ARE, and determines whether or not there are resources (the N reflecting elements 21 that are available and the maximum achievable rate AR _max ) to be used for future communications.

When the transmitter 1 determines that there are resources available for future transmission, it sends a communication notification to the control means 223 of the IRS 2 and begins communication.

After receiving a communication notification from the transmitter 1, the control means 223 of the IRS 2 controls the pattern of the N reflecting elements 21 to a reflection pattern when the reflected beam forming vector φ _max is used, based on the reflected beam forming vector φ _max that it holds.

Therefore, by estimating the usable reflecting elements among the N reflecting elements 21 and the maximum achievable rate AR _max among the _E achievable rates AR ₁ to ARE, resources for communication using the N reflecting elements 21 can be secured based on the usable reflecting elements among the N reflecting elements 21 and the maximum achievable rate AR _max , and therefore it is possible to create a schedule for using the reflecting elements by determining in advance whether enough reflecting elements are actually available to ensure a sufficient transmission rate.

Furthermore, the estimation means 43 inputs the time series data B/I ₁ (t) to B/I _N (t) to the probabilistic neural networks PNN ^SLT _1 to PNN ^SLT _N, respectively, and the probabilistic neural networks PNN ^SLT _1 to PNN ^SLT _N execute the flowchart shown in FIG. 12 in parallel to output the availability prediction results U ₁ (t) to U _N (t) to the estimation means 43, which means that the estimation means 43 inputs the first idle state/busy state data to one first machine learning device (PNN ^SLT _1 to PNN ^{SLT _N)} . _N), and executes a first process of receiving from one first machine learning device a usability prediction result corresponding to the maximum probability among the probabilities that a first class classified according to the idle state of the first idle state/busy state data matches a first learning stage class, which is a class classified according to the idle state of the first idle state/busy state data of the learning stage acquired in the learning stage, and executes the same process as the first process for the second idle state/busy state data to the Nth idle state/busy state data instead of the first idle state/busy state data in the first process, and for all of the (N-1) first machine learning devices, thereby estimating the usability prediction results for the N reflective elements.

Furthermore, the estimation means 43 inputs the achievable rates AR ₁ (t)_ _COL to _ARE (t)_ _COL to the probabilistic neural networks PNN ^AR — 1 to PNN ^AR — E, respectively, and the probabilistic neural networks PNN ^AR — 1 to PNN ^AR — E execute the flowchart shown in FIG. 13 in parallel to output the achievable rates ^AR ₁ (t)_ _PRD to _ARE (t) ^_ _PRD to the estimation means 43, respectively. _E), and receiving from one of the first machine learning machines PNN AR — 1 to PNN AR _N an achievable rate corresponding to the maximum probability among the probabilities that a first class classified according to the achievable rate of the first achievable rate data matches a second learning-stage class ^that is a class classified according to the achievable rate of the first achievable rate data of the learning stage acquired in the learning stage, and performing the same process as the second process for the second achievable rate data to the Nth achievable rate data and all of the ( ^E −1) second machine learning machines, instead of the first achievable rate data in the second process.

In this case, AR ₁ (t) — _COL to AR _E (t) — _COL shown in Figure 9 are respectively associated with V reflected beamforming vectors φ ₁ to φ _V of N reflecting elements 21 (each of which is a reflected beamforming vector φ _m when one or more reflecting elements 21 are used simultaneously), and each constitutes "first achievable rate data" to "Eth achievable rate data" in which the achievable rates are arranged in chronological order.

In an embodiment of the present invention, the operation of the estimation device 4 may be realized by software. In this case, the estimation device 4 includes a CPU (Central Processing Unit), a ROM (Read Only Memory), and a RAM (Random Access Memory). The ROM stores a program Prog_A consisting of the steps of the flowchart shown in Figure 11 (including the flowcharts shown in Figures 12 and 13).

The CPU reads out the program Prog_A from the ROM and executes the read out program Prog_A to estimate usable reflecting elements among the N reflecting elements 21 and E achievable rates AR ₁ to _ARE , and obtain a reflected beamforming vector φ _max of the N reflecting elements 21. The RAM temporarily stores a codebook including the reflected beamforming vector φ of the N reflecting elements 21.

Furthermore, the program Prog_A may be recorded on a recording medium such as a CD, a DVD, etc. and distributed. When the recording medium on which the program Prog_A is recorded is attached to a computer, the computer reads and executes the program Prog_A from the recording medium to estimate usable reflecting elements among the N reflecting elements 21 and E achievable rates AR ₁ to _ARE , and obtains the reflected beamforming vector φ _max of the N reflecting elements 21.

Therefore, the recording medium on which the program Prog_A is recorded is a computer-readable recording medium.

Figure 14 is a schematic diagram of another wireless communication system according to an embodiment of the present invention. The wireless communication system according to an embodiment of the present invention may be wireless communication system 10A shown in Figure 14.

Referring to FIG. 14, the wireless communication system 10A includes a transmitter 1, IRSs 20 and 30, and a receiver 3.

Transmitter 1, IRS 20, 30, and receiver 3 are arranged in a wireless communication space. Transmitter 1, IRS 20, 30, and receiver 3 are arranged in a position where the arrangement [Transmitter 1 - IRS 20, 30 - Receiver 3] is not on a straight line. More specifically, transmitter 1, IRS 20, and receiver 3 are arranged in a position where the arrangement [Transmitter 1 - IRS 20 - Receiver 3] is not on a straight line, and transmitter 1, IRS 30, and receiver 3 are arranged in a position where the arrangement [Transmitter 1 - IRS 30 - Receiver 3] is not on a straight line.

Then, when the line connecting the transmitter 1 and the IRS 20 is L3 and the line connecting the IRS 20 and the receiver 3 is L4, the transmitter 1, the IRS 20, and the receiver 3 are arranged so that the angle _α1 formed by the lines L3 and L4 is greater than 0. Furthermore, when the line connecting the transmitter 1 and the IRS 30 is L5 and the line connecting the IRS 30 and the receiver 3 is L6, the transmitter 1, the IRS 30, and the receiver 3 are arranged so that the angle _α2 formed by the lines L5 and L6 is greater than 0.

In other words, transmitter 1, IRS 20, and receiver 3 are arranged non-linearly in the wireless communication space, and transmitter 1, IRS 23, and receiver 3 are arranged non-linearly in the wireless communication space. As a result, transmitter 1, IRS 20, 30, and receiver 3 are arranged non-linearly in the wireless communication space.

Each of IRSs 20 and 30 has the same configuration as IRS 2 described above. The estimation device 4 of each of IRSs 20 and 30 receives the received signal y(t) from the receiver 3 and estimates the frequency response FQR_0 of the propagation path from the transmitter 1 to the receiver 3 using the same method as the estimation device 4 of IRS 2.

In addition, the estimation device 4 of the IRS 20 estimates usable reflecting elements among the N reflecting elements 21 of the IRS 20 and _E achievable rates AR ₁ to ARE according to the flowchart shown in FIG. 11 (including the flowcharts shown in FIGS. 12 and 13), and obtains the reflected beamforming vector φ _max of the N reflecting elements 21 of the IRS 20.

Furthermore, the estimation device 4 of the IRS 30 estimates usable reflecting elements among the N reflecting elements 21 of the IRS 30 and E achievable rates AR ₁ to _ARE according to the flowchart shown in FIG. 11 (including the flowcharts shown in FIGS. 12 and 13), and obtains the reflected beamforming vector φ _max of the N reflecting elements 21 of the IRS 30.

The IRS controller 22 of IRS20 is connected to the IRS controller 22 of IRS30 via optical fiber, and can communicate with the IRS controller 22 of IRS30.

If the communication capacity for wireless communication from the transmitter 1 to the receiver 3 is insufficient using only the N reflecting elements 21 of the IRS 20, the IRS controller 22 of the IRS 20 estimates the amount of resources RSC available for wireless communication using the N reflecting elements 21 of the IRS 20 and the N reflecting elements 21 of the IRS 30.

Here, the frequency response matrix H(t) of the propagation path estimated by the estimation means 43 of IRS20 using equations (1) to (6) is defined as frequency response matrix H ₂₀ (t), and the frequency response matrix H(t) of the propagation path estimated by the estimation means 43 of IRS30 using equations (1) to (6) is defined as frequency response matrix H ₃₀ (t).

The IRS controller 22 of the IRS 20 instructs the IRS controller 22 of the IRS 30 to transmit the frequency response matrix H ₃₀ (t) to itself.

In response to an instruction from the IRS controller 22 of the IRS 20 , the IRS controller 22 of the IRS 30 transmits the frequency response matrix H ₃₀ (t) in the IRS 30 to the IRS controller 22 of the IRS 20 .

The IRS controller 22 of the IRS 20 receives the frequency response matrix H ₃₀ (t) from the IRS controller 22 of the IRS 30 .

Then, in the IRS controller 22 of the IRS 20 , the estimation means 43 of the estimation device 4 receives the frequency response matrix H ₃₀ (t) from the communication means 222 .

The control means 223 of the IRS 20 holds a codebook CD_B containing a reflection beamforming vector φ that controls the reflection coefficient of the reflection element 21.

The codebook CD_B is configured by arranging the reflected beamforming vector φ in a matrix. The codebook CD_B consists of 2N column elements and, for example, 8 to 16 row elements. Here, the codebook CD_B is assumed to include 8 row elements.

When the IRS controller 22 of IRS 20 uses N reflecting elements 21 of IRS 20 and N reflecting elements 21 of IRS 30, the IRS controller 22 of IRS 20 controls 2N reflecting elements 21. The codebook CD_B then includes 8 x 2 reflected beamforming vectors φ.

The control means 223 of the IRS 20 selects patterns REF_ptn_1 to ^REF_ptn_E2 of ^E2 reflective elements from the codebook CD_B when one or more reflective elements 21 of the 2N reflective elements are used simultaneously.

When collecting time series data TSD_1(t)_20, the collection means 41 of the IRS 20 receives from the control means 223 of the IRS 20 a period ON_period/period OFF_period_n(t) in which a period OFF_period is a period during which one or more of the 2N reflective elements 21 are turned OFF, and a period ON_period is a period during which reflective elements other than one or more of the 2N reflective elements 21 are turned ON, arranged in time series.

Then, the collection means 41 associates "I" with one slot SLT based on the period OFF_period, and associates "B" with one slot SLT based on the period ON_period, and continues in the same manner to generate time series data TSD_1(t)_20.

In this case, the collection means 41 detects that the reflective elements that are turned on (reflective elements other than one or more reflective elements 21) in one slot SLT are in a busy state B, and detects that the reflective elements that are turned off (reflective elements 21) in one slot SLT are in an idle state I, and so on, generating time series data TSD_1(t)_20 in the same manner.

Then, the collecting means 41 stores the generated time series data TSD_1(t)_20 in the storage means 42. The time series data TSD_1(t)_20 is made up of time series data B/I ₁ (t) to B/I _2N (t).

The collection means 41 of the IRS 20 calculates the achievable rate R using equations (7) to (10) for all of the patterns REF_ptn_1 to ^REF_ptn_E2 of the ^E2 reflective elements to collect time series data TSD_2(t)_20.

The time series data TSD_2(t)_20 consists of AR ₁ (t) _{_COL} to AR _E^2 (t) _{_COL} , where "E^2" means E squared.

Figure 15 is a schematic diagram of the learners 44 and 45 when the IRS controller 22 of the IRS 20 controls 2N reflecting elements 21.

15, when the IRS controller 22 of the IRS 20 controls 2N reflecting elements 21, the learning device 44 is made up of 2N probabilistic neural networks PNN ^SLT — 1 to PNN ^{SLT —} 2N (see FIG. 15(a)), and the learning device 45 is made up of ^E2 probabilistic neural networks PNN ^AR — 1 to PNN ^AR — E ² (see FIG. 15(b)).

Each of the probabilistic neural networks PNN ^{SLT —} 1 to PNN ^SLT — 2N has the same configuration as the probabilistic neural network PNN ^SLT — 1 shown in FIG.

When time series data B/I _e* (t) (any of time series data B/I ₁ (t) to B/I _2N (t)) is input to the probabilistic neural networks PNN ^SLT _1 to ^{PNN SLT} _{_e} ^{* (} e* is any of 1 to 2N), the probabilistic neural networks output the class c max (i.e., the usability prediction result U _{e*(t)) when the maximum probability is obtained that the input time series data B/I e* (} t) matches the class c _learn of the data (time series data TSD_1_learn(t)) stored in the prediction matrix P _{SLT r} (the prediction matrix held in the summation layer).

Therefore, when time series data B/I ₁ (t) to B/I _2N (t) are input to the probabilistic neural networks PNN ^SLT _ 1 to PNN ^SLT _ 2N, the probabilistic neural networks output classes c _{max — 1} to c max — _2N (i.e., usability prediction results U ₁ (t) to U _2N (t)) that provide the maximum probability (see (a) of FIG. 15).

6. Each of the E ² probabilistic neural networks PNN ^AR — 1 to PNN ^AR — E ² has the same configuration as the probabilistic neural network PNN ^SLT — 1 shown in FIG.

When time series data AR _e* (t) _{_COL} is input, the probabilistic neural networks PNN ^AR _1 to PNN ^AR _e ^* (e ^* is any of 1 to ^E2 ) output a positive integer (a positive integer indicating the achievable rate ^AR ₎ contained in the row T that has received the maximum probability out of the n _R probabilities received from the n _R rows of the prediction matrix P AR r.

Therefore, when the time series data AR ₁ (t) _{_COL} to AR E _^2 (t) _{_COL} are input to the probabilistic neural networks PNN ^AR — 1 to PNN ^AR — E ² , respectively, they output a positive integer contained in the row T that has received the maximum probability (i.e., a positive integer indicating the achievable rate AR ₁ (t) _{_PRD} to AR _E^2 (t) _{_PRD} ).

In the IRS 20, the estimation means 43 receives from the learner 44 usable prediction results indicating usable reflective elements among the 2N reflective elements, and receives from the learner 45 achievable rates AR ₁ (t) _{_PRD} to AR _E^2 (t) _{_PRD} of the E ² reflective elements.

Then, in the IRS 20, the estimation means 43 detects the maximum achievable rate AR _max among the achievable rates AR ₁ (t) — _PRD to AR _E^2 (t) — _PRD of the E ^two reflecting elements, and detects the reflected beamforming vector φ _max from the codebook CD_B when the detected maximum achievable rate AR _max is obtained.

Then, in the IRS 20, the estimation means 43 creates a communication schedule using the available reflecting elements out of 2N, the maximum achievable rate AR _max and the reflected beamforming vector φ _max .

Then, in the IRS 20, the estimation means 43 outputs the usable reflecting elements out of the 2N reflecting elements, the maximum achievable rate AR _max , the reflected beamforming vector φ _max and the communication schedule to the control means 223.

In the IRS 20 , the control means 223 receives the available reflecting elements out of 2N reflecting elements, the maximum achievable rate AR _max , the reflected beamforming vector φ _max and the communication schedule from the estimation means 43 .

Then, in the IRS 20, the control means 223 holds the reflected beamforming vector φ _max and transmits to the transmitter 1 the resource amount RSC_20,30 indicating the usable reflecting elements out of the 2N reflecting elements 21 and the maximum achievable rate AR _max , as well as a communication schedule.

The transmitter 1 receives the resource amount RSC_20, 30 and the communication schedule from the IRS controller 22 of the IRS 20.

Then, when the transmitter 1 starts wireless communication with the receiver 3, it sends a communication notification to the IRS controller 22 of the IRS 20.

The IRS controller 22 of the IRS 20 receives a communication notification from the transmitter 1, and then when wireless communication starts, controls the reflection pattern of the 2N reflecting elements 21 in the IRSs 20 and 30 based on the reflection beamforming vector φ _max in the IRSs 20 and 30.

Furthermore, in an embodiment of the present invention, the IRS controller 22 of IRS 30 may control 2N reflecting elements in IRSs 20 and 30 in the same manner as the IRS controller 22 of IRS 20.

The operation of the estimation device 4 of the IRS 20 is performed according to the flowchart shown in Fig. 11 (including the flowcharts shown in Figs. 12 and 13). In this case, "N reflecting elements" should be read as "2N reflecting elements,""N usable reflecting elements" should be read as "2N usable reflecting elements,""E achievable rates R ₁ to R _E " should be read as "E ² achievable rates R ₁ to R _E^2 ."

Figure 16 is a schematic diagram showing one specific example of the wireless communication system shown in Figure 14. Referring to Figure 16, sidewalks are located on both sides of the road.

IRS-1a, IRS-2a, and IRS-3a will be placed facing the sidewalk on one side of the road. IRS-1a, IRS-2a, and IRS-3a will be fixed to the side of the building.

IRS-1b, IRS-2b, and IRS-3b will be placed facing the sidewalk on the other side of the road. IRS-1b, IRS-2b, and IRS-3b will be fixed to the side of the building.

IRS controller 1a is connected to IRS controller 1b and IRS controller 2a via optical fiber, and is also connected to IRS-1a and IRS-2a.

IRS controller 2a is connected to IRS controller 1a and IRS controller 2b via optical fiber, and is also connected to IRS-2a and IRS-3a.

IRS controller 1b is connected to IRS controller 1a and IRS controller 2b via optical fiber, and is also connected to IRS-1b and IRS-2b.

IRS controller 2b is interconnected with IRS controller 2a and IRS controller 1b via optical fiber, and is also connected to IRS-2b and IRS-3b.

In FIG. 16, IRS controller 1a, IRS controller 2a, IRS controller 1b, and IRS controller 2b each control N reflecting elements 21 of two IRSs.

The mobile terminal then performs wireless communication using, for example, IRS-2a, IRS-2b, and IRS-3b.

In this case, before wireless communication by the mobile terminal is started, the estimation device 4 of the IRS controller 2a estimates usable reflecting elements among the N reflecting elements 21 of the IRS-2a and the achievable rate AR according to the flowchart shown in FIG. 11 (including the flowcharts shown in FIGS. 12 and 13), and obtains the reflected beamforming vector φ _max of the N reflecting elements 21 when the maximum achievable rate AR _max is obtained in the IRS-2a.

Then, the estimation device 4 of the IRS controller 2a creates a communication schedule based on the available reflecting elements among the N reflecting elements 21 of the IRS-2a, the maximum achievable rate AR _max and the reflecting beamforming vectors φ _max of the N reflecting elements 21.

The estimation unit 4 of the IRS controller 2a then transmits to the mobile terminal the usable reflecting elements among the N reflecting elements 21 of the IRS-2a, the maximum achievable rate AR _max and the communication schedule.

On the other hand, before wireless communication by the mobile terminal is started, the estimation device 4 of the IRS controller 2b estimates usable reflecting elements among the N reflecting elements 21 of the IRS-2b and the maximum achievable rate AR _max according to the flowchart shown in FIG. 11 (including the flowcharts shown in FIGS. 12 and 13), and obtains the reflected beamforming vector φ _max of the N reflecting elements 21 of the IRS-2b.

Furthermore, before wireless communication by the mobile terminal is started, the estimation device 4 of the IRS controller 2b estimates usable reflecting elements among the N reflecting elements 21 of the IRS-3b and the maximum achievable rate AR _max according to the flowchart shown in FIG. 11 (including the flowcharts shown in FIGS. 12 and 13), and obtains the reflected beamforming vector φ _max of the N reflecting elements 21 of the IRS-3b.

Then, the estimation device 4 of the IRS controller 2b creates a communication schedule based on the usable reflecting elements among the N reflecting elements 21 of the IRS-2b, the maximum achievable rate AR _max , and the reflected beamforming vector φ _max of the N reflecting elements 21, and based on the usable reflecting elements among the N reflecting elements 21 of the IRS-3b, the maximum achievable rate AR _max , and the reflected beamforming vector φ _max of the N reflecting elements 21.

Then, the estimation device 4 of the IRS controller 2b transmits to the mobile terminal the usable reflecting elements among the N reflecting elements 21 of the IRS-2b and the maximum achievable rate AR _max , the usable reflecting elements among the N reflecting elements 21 of the IRS-3b and the maximum achievable rate AR _max , and a communication schedule.

Then, IRS controller 2a and IRS controller 2b receive a communication notification from the mobile terminal. Then, when wireless communication by the mobile terminal is started, IRS controller 2a controls the reflection pattern of the N reflecting elements of IRS-2a in accordance with the reflection beamforming vector φ _max of IRS-2a, and IRS controller 2b controls the reflection pattern of the N reflecting elements of IRS-2b in accordance with the reflection beamforming vector φ _max of IRS-2b, and also controls the reflection pattern of the N reflecting elements of IRS-3b in accordance with the reflection beamforming vector φ _max of IRS-3b.

This allows a mobile terminal to use three IRS-2a, IRS-2b, and IRS-3b in a wireless communication space where buildings are located on both sides of a road, allowing wireless communication while minimizing the reduction in radio wave reach caused by buildings.

In one specific example of the wireless communication system shown in FIG. 16 , if the estimation device 4 of the IRS controller 2 a estimates usable reflecting elements among the N reflecting elements of IRS-2 a and the maximum achievable rate AR _max of the N reflecting elements, and the estimation device 4 of the IRS controller 2 b estimates usable reflecting elements among the N reflecting elements of IRS-2 b and the maximum achievable rate AR _max of the N reflecting elements, and also estimates usable reflecting elements among the N reflecting elements of IRS-3 b, then the estimated usable reflecting elements and the maximum achievable _{rate AR max of} the N reflecting elements of IRS-2 a, the estimated usable reflecting elements and the maximum achievable rate AR _{max of the N reflecting elements of IRS-2 b, and the estimated maximum achievable rate AR max} of the usable reflecting elements and the N reflecting elements of IRS-3 b can be calculated. _max , communication resources can be secured using the N reflecting elements of IRS-2a, the N reflecting elements of IRS-2b, and the N reflecting elements of IRS-3b, so that a schedule for using the reflecting elements can be created by determining in advance whether enough reflecting elements are actually available to ensure a sufficient transmission rate.

According to the above-described embodiment, the estimation device according to the embodiment of the present invention comprises:
An estimation device for estimating information useful for predicting a usage schedule, which is a schedule for using N reflecting elements, in a wireless communication system in which a transmitter having M _T (M _T is an integer of 2 or more) antennas, a first IRS which is an electromagnetic wave reflector having N (N is an integer of 2 or more) reflecting elements whose reflection characteristics can be dynamically changed, and a receiver having M _R (M _R is an integer of 2 or more) antennas are non-linearly arranged, comprising:
a collecting means for collecting first time series data indicating, for each of the N reflecting elements, time series data consisting of a busy state in which the reflecting element is in use and an idle state in which the reflecting element is not in use, and second time series data indicating an achievable rate which is an achievable transmission rate of the N reflecting elements when a transmitter is transmitting a signal using the M _T antennas;
The system may include an estimation means that performs a first estimation process that uses a first learning device to estimate a usability prediction result for each of N reflecting elements, which is the result of predicting whether or not the reflecting elements in the first IRS are usable, based on first time series data, and a second estimation process that uses a second learning device to estimate the achievable rates of the N reflecting elements of the first IRS in the propagation path for those reflecting elements that are in use among the N reflecting elements, based on second time series data.

If the estimation means estimates the predicted availability results of N of the N reflecting elements and the achievable rate of the reflecting elements that are being used among the N reflecting elements, resources for communication using the N reflecting elements can be secured based on the estimated predicted availability results of the N reflecting elements and the achievable rate, so that a schedule for using the reflecting elements can be created by determining in advance whether enough reflecting elements are actually available to ensure a sufficient transmission rate.

Furthermore, the program according to the embodiment of the present invention is
A program for causing a computer to execute estimation of information useful for predicting a usage schedule, which is a schedule for using N reflecting elements, in a wireless communication system in which a transmitter having M _T (M _T is an integer of 2 or more) antennas, a first IRS which is an electromagnetic wave reflector having N (N is an integer of 2 or more) reflecting elements whose reflection characteristics can be dynamically changed, and a receiver having M _R (M _R is an integer of 2 or more) antennas are non-linearly arranged, the program comprising:
a first step in which a collecting means collects first time series data for each of the N reflecting elements, the first time series data being composed of a busy state in which the reflecting element is in use and an idle state in which the reflecting element is not in use, and second time series data being indicative of an achievable transmission rate of the N reflecting elements when a transmitter is transmitting a signal using the M _T antennas;
The estimation means simply causes the computer to execute a second step in which the estimation means executes a first estimation process in which, based on first time series data, a usability prediction result is estimated for each of N reflecting elements, which is the result of predicting whether or not the reflecting elements in the first IRS are usable using a first learning device, and a second estimation process in which, based on second time series data, a second learning device is used to estimate the achievable rates of the N reflecting elements of the first IRS in the propagation path for those reflecting elements that are in use among the N reflecting elements.

If the program causes a computer to estimate the predicted availability of N of the N reflecting elements and the achievable rate for the reflecting elements that are in use among the N reflecting elements, resources for communication using the N reflecting elements can be secured based on the estimated predicted availability of the N reflecting elements and the achievable rate for the reflecting elements that are in use among the N reflecting elements. This makes it possible to create a schedule for using the reflecting elements by determining in advance whether enough reflecting elements are actually available to ensure a sufficient transmission rate.

In an embodiment of the present invention, the N usability prediction results of the N reflecting elements 21 and the achievable rates R ₁ to R _N of the reflecting elements being used among the N reflecting elements constitute "information useful for predicting a usage schedule which is a schedule for using the N reflecting elements 21." Note that the reflected beamforming vectors φ _max of the N reflecting elements 21 may be included in "information useful for predicting a usage schedule which is a schedule for using the N reflecting elements 21."

Furthermore, in this embodiment of the present invention, learning device 44 constitutes a "first learning device," and learning device 45 constitutes a "second learning device."

Furthermore, in this embodiment of the present invention, the probabilistic neural networks PNN ^SLT — 1 to PNN ^SLT — N constitute "N first machine learning machines," and the probabilistic neural networks PNN ^AR — 1 to PNN ^AR _E constitute "E second machine learning machines."

Furthermore, in this embodiment of the present invention, the time series data B/I ₁ (t) constitutes the "first idle state/busy state data," and the time series data B/I _N (t) constitutes the "Nth idle state/busy state data."

The embodiments disclosed herein should be considered in all respects to be illustrative and not restrictive. The scope of the present invention is indicated by the claims, not by the description of the above embodiments, and is intended to include all modifications within the meaning and scope of the claims.

This invention applies to an estimation device, a wireless communication system including the same, and a program for causing a computer to execute the device.

1 Transmitter, 2, 20, 30 IRS, 3 Receiver, 4 Estimation device, 10, 10A Wireless communication system, 21 Reflection element, 22 IRS controller, 41 Collection means, 42 Storage means, 43 Estimation means, 44, 45 Learner, 221 Antenna, 222 Communication means, 223 Control means.

Claims (16)

An estimation device for estimating information useful for predicting a usage schedule, which is a schedule for using N reflecting elements, in a wireless communication system in which a transmitter having M _T antennas (M _T is an integer of 2 or more), a first IRS which is an electromagnetic wave reflector having N reflecting elements ( _N is an integer of 2 or more) whose reflection characteristics can be dynamically changed, and a receiver having M _R antennas are non-linearly arranged, the device comprising:
a collecting means for collecting first time series data indicating, for each of the N reflecting elements, time series data consisting of a busy state in which the reflecting element is in use and an idle state in which the reflecting element is not in use, and second time series data indicating an achievable rate, which is an achievable transmission rate of the N reflecting elements when the transmitter is transmitting a signal using the M _T antennas;
An estimation device comprising an estimation means that performs a first estimation process that uses a first learning device to estimate a usability prediction result for each of the N reflecting elements, which is a result of predicting whether the reflecting elements in the first IRS are usable or not, based on the first time series data, and a second estimation process that uses a second learning device to estimate the achievable rate of the N reflecting elements of the first IRS for used reflecting elements among the N reflecting elements in a propagation path in which the transmitter, the first IRS, and the receiver are arranged non-linearly, based on the second time series data. The estimation device of claim 1, wherein the estimation means, in the first estimation process, inputs the first time series data to the first learning device and estimates usable reflecting elements among the N reflecting elements by receiving from the first learning device the idle state corresponding to the maximum probability among probabilities that a first class classified according to the idle state of the first time series data matches a first learning stage class, which is a class classified according to the idle state of the first learning time series data acquired in a learning stage; and, in the second estimation process, inputs the second time series data to the second learning device and estimates the achievable rate of a used reflecting element among the N reflecting elements by receiving from the second learning device the achievable rate corresponding to the maximum probability among probabilities that a second class classified according to the achievable rate of the second time series data matches a second learning stage class, which is a class classified according to the achievable rate of the second learning time series data acquired in a learning stage. the first learning device comprises N first machine learning devices, the second learning device comprises E second machine learning devices , the number of which is the number of reflection patterns of one or more of the N reflection elements when the one or more reflection elements are simultaneously used, the first time series data comprises first idle state/busy state data to Nth idle state/busy state data, each of which arranges the idle state and the busy state in time series, the second time series data corresponds to E reflection patterns and comprises first achievable rate data to Eth achievable rate data, each of which arranges the achievable rates in time series,
The estimation means
In the first estimation process, the first idle state/busy state data is input to one first machine learning machine among the N first machine learning machines , and a first process is performed to receive from the one first machine learning machine the idle state corresponding to the maximum probability among probabilities that the first class classified according to the idle state of the first idle state/busy state data matches the first learning stage class, which is the class classified according to the idle state of the first idle state/busy state data of the learning stage acquired in a learning stage, and the same process as the first process is performed for second idle state/busy state data to Nth idle state/busy state data among the first idle state/busy state data to Nth idle state/busy state data instead of the first idle state/busy state data in the first process, and for all of the (N-1) first machine learning machines, thereby estimating the N usability prediction results of the N reflecting elements;
3. The estimation device according to claim 2, wherein, in the second estimation process, the first achievable rate data is input to one of the second machine learning machines, and a second process is performed to receive from the one first machine learning machine the achievable rate corresponding to the largest probability among probabilities that the first class classified according to the achievable rate of the first achievable rate data matches the second learning stage class, which is a class classified according to the achievable rate of the first achievable rate data of the learning stage acquired in a learning stage, and the achievable rate of a used reflecting element among the N reflecting elements is estimated by performing the same process as the second process for second achievable rate data through E achievable rate data among the first achievable rate data through E achievable rate data and all of the (E−1) second machine learning machines , instead of the first achievable rate data in the second process. the N first machine learning machines execute the first process in parallel in the first estimation process;
The estimation device according to claim 3 , wherein the E second machine learning machines execute the second process in parallel in the second estimation process. each of the N first machine learning machines classifies the idle states acquired in a first learning stage into a first plurality of classes and stores the classes in a first prediction matrix, and when one of the idle state/busy state data is input, selects one of the classes from the first plurality of classes stored in the first prediction matrix that has the highest probability of matching the idle state of the input one of the idle state/busy state data, and outputs the availability prediction result to the estimation means based on the selection of the one of the classes;
5. The estimation device according to claim 3, wherein each of the E second machine learning machines classifies the achievable rates obtained in the second learning stage into a second plurality of classes and stores them in a second prediction matrix, and when one piece of achievable rate data is input , selects the one class from the second plurality of classes that has the highest probability of matching the achievable rate of the input piece of achievable rate data, and outputs the achievable rate classified into the one selected class to the estimation means. 6. The estimation device according to claim 1, wherein in the second estimation process, the estimation means converts propagation paths to the M _R antennas of the receiver into M _R independent propagation paths, calculates M _R transmission rates for the M _R independent propagation paths, and sets the sum of the calculated M _R transmission rates as the achievable rate. 7. The estimation device according to claim 6, wherein said estimation means calculates said M _R transmission rates based on signal-to-interference-plus-noise power ratios of said M _R antennas in said second estimation process. A radio communication system comprising: a first estimating device comprising the estimating device according to any one of claims 1 to 7; the transmitter; the receiver; and the first IRS;
The first estimator is located within the first IRS. a second estimation device comprising the estimation device according to any one of claims 1 to 7; and a second IRS which is an electromagnetic wave reflector having N ^* (N ^* is an integer of 2 or more) reflecting elements whose reflection characteristics can be dynamically changed;
The wireless communication system of claim 8 , wherein the second estimator is located within the second IRS. A program for causing a computer to execute estimation of information useful for predicting a usage schedule, which is a schedule for using N reflecting elements, in a wireless communication system in which a transmitter having M _T (M _T is an integer of 2 or more) antennas, a first IRS which is an electromagnetic wave reflector having N (N is an integer of 2 or more) reflecting elements whose reflection characteristics can be dynamically changed, and a receiver having M _R (M _R is an integer of 2 or more) antennas are non-linearly arranged, the program comprising:
a first step of collecting first time series data for each of the N reflecting elements, the first time series data being composed of a busy state in which the reflecting element is in use and an idle state in which the reflecting element is not in use, and second time series data being indicative of an achievable rate, which is an achievable transmission rate of the N reflecting elements when the transmitter is transmitting a signal using the M _T antennas;
A program for causing a computer to execute a first estimation process in which a usability prediction result, which is a result of predicting whether the reflecting element in the first IRS is usable or not, is estimated for each of the N reflecting elements using a first learning device based on the first time series data, and a second estimation process in which, based on the second time series data, a second learning device is used to estimate the achievable rate of the N reflecting elements of the first IRS for used reflecting elements among the N reflecting elements in a propagation path in which the transmitter, the first IRS, and the receiver are arranged non-linearly. 11. The program for causing a computer to execute the program according to claim 10, wherein in the first estimation process of the second step, the first time series data is input to the first learning device, and the idle state corresponding to the maximum probability of a first class classified according to the idle state of the first time series data matching a first learning stage class that is a class classified according to the idle state of first learning time series data acquired in a learning stage is output from the first learning device, thereby estimating usable reflecting elements among the N reflecting elements; and in the second estimation process of the second step, the second time series data is input to the second learning device, and the achievable rate corresponding to the maximum probability of a second class classified according to the achievable rate of the second time series data matching a second learning stage class that is a class classified according to the achievable rate of second learning time series data acquired in a learning stage is output from the second learning device, thereby estimating an achievable rate of a used reflecting element among the N reflecting elements. the first learning device comprises N first machine learning devices, the second learning device comprises E second machine learning devices , the number of which is the number of reflection patterns of one or more of the N reflection elements when the one or more reflection elements are simultaneously used, the first time series data comprises first idle state/busy state data to Nth idle state/busy state data, each of which arranges the idle state and the busy state in time series, the second time series data corresponds to E reflection patterns and comprises first achievable rate data to Eth achievable rate data, each of which arranges the achievable rates in time series,
In the first estimation process of the second step, the first idle state/busy state data is input to one first machine learning machine among N first machine learning machines , and a first process is executed in which the idle state corresponding to the maximum probability among probabilities that the first class classified according to the idle state of the first idle state/busy state data matches the first learning stage class, which is the class classified according to the idle state of the first idle state/busy state data of the learning stage acquired in a learning stage , is output from the one first machine learning machine, and the same process as the first process is executed for second idle state/busy state data to Nth idle state/busy state data among the first idle state/busy state data to Nth idle state/busy state data instead of the first idle state/busy state data in the first process, and for all of the (N-1) first machine learning machines , thereby estimating the N usability prediction results of the N reflecting elements,
12. The program for causing a computer to execute the program according to claim 11, wherein in the second estimation process of the second step, the first achievable rate data is input to one of the second machine learning machines, and a second process is executed in which the achievable rate corresponding to the largest probability of the first class classified according to the achievable rate of the first achievable rate data matching the second learning stage class, which is a class classified according to the achievable rate of the first achievable rate data of the learning stage acquired in a learning stage, is output from the one first machine learning machine , and ... of the N reflective elements used is estimated by executing the same process as the second process for second achievable rate data to E achievable rate data among the first achievable rate data to E achievable rate data and all of the (E-1) second machine learning machines , instead of the first achievable rate data in the second process. In the first estimation process of the second step, the first process is executed in parallel,
The program for causing a computer to execute the program according to claim 12 , wherein the second estimation process in the second step is executed in parallel with the second process. the idle states acquired in the first learning stage are classified into a first plurality of classes and stored in a first prediction matrix, and when one of the idle state/busy state data is input, one of the classes having the highest probability of matching the idle state of the input one of the idle state/busy state data is selected from the first plurality of classes stored in the first prediction matrix, and the availability prediction result is output based on the selection of the one of the class;
14. A program for causing a computer to execute the program described in claim 12 or 13, wherein the achievable rates obtained in the second learning stage are classified into a second plurality of classes and stored in a second prediction matrix, and when one of the achievable rate data is input, the one of the classes that has the highest probability of matching the achievable rate of the input one of the achievable rate data is selected from the second plurality of classes, and the achievable rate classified into the selected one of the classes is output . 15. The program executed by a computer according to claim 10, wherein in the second estimation process of the second step, propagation paths to the M _R antennas of the receiver are converted into M _R independent propagation paths, M _R transmission rates for the M _R independent propagation paths are calculated , and a sum of the calculated M _R transmission rates is set as the achievable rate. 16. The program executed by a computer according to claim 15, wherein in the second estimation process of the second step, the M _R transmission rates are calculated based on signal-to-interference-plus-noise power ratios of the M _R antennas.