JP2017182528A - Parameter adjustment device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a parameter adjustment device with which it is possible to shorten a computation time for searching for the value of a parameter set so that satisfies required performance.SOLUTION: Provided is a parameter adjustment device for adjusting the value of a parameter set for a first group intelligent search, said device including: a first stage processing unit (S4) for determining a capability value for as many parameter sets as the number of parameter sets for low-density test, and extracting a parameter set candidate applying a second group intelligent search; a second stage processing unit (S5) for executing a high-density test on a second stage evaluation parameter set determined on the basis of the parameter set candidate, for evaluating capability by as many test cases as the number of high-density test cases; and a parameter determination unit (S5, S7) for determining a second stage evaluation parameter set the evaluation result of which, by the second stage processing unit, satisfies the required performance as a parameter set used in a search of solution. In the first stage processing unit, a low-density test is executed that determines the capability value of a parameter set by as many test cases as the number of low-density test cases.SELECTED DRAWING: Figure 6

Description

  The present invention relates to a parameter adjustment device, and more particularly to a technique for shortening a calculation time for adjusting a parameter to satisfy a required performance in a group intelligence search.

  A technique for searching for an optimum value by swarm intelligence search is widely known. Also, a technique disclosed in Patent Document 1 is known as a technique for searching for an optimum value.

  Patent Document 1 discloses a technique for estimating the arrival direction of radio waves. In Patent Document 1, the direction of radio wave arrival is estimated by the Pseudo-doppler method. In the Pseudo-doppler method, an antenna is mounted on a rotating plate or the like, and the antenna is circularly moved to create a Doppler shift for the radio wave emitted from the radio wave source.

  The created Doppler shift is maximized on the plus side or the minus side when the antenna velocity vector is in the direction toward the radio wave source and when the antenna speed vector is in the direction opposite to the radio wave source. . This change in Doppler shift is expressed by a model including the direction of arrival of radio waves as a variable. Then, the optimum value of the radio wave arrival direction is determined by searching for a model that most closely matches the observed signal. And this optimal value is estimated as a radio wave arrival direction.

Japanese Patent Laid-Open No. 2006-8913

  The model of Patent Document 1 has three variables: phase, amplitude, and radio wave arrival direction. If the incoming radio wave is only one wave and the required angular resolution is 10 degrees, the incoming direction of the radio wave is only 36 ways. However, multiple arrival directions of radio waves may exist due to the influence of multipath. If radio waves arrive from four directions due to the influence of multipath, there are 36 4 combinations. Therefore, if all combinations are searched, the calculation time for searching becomes too long.

  Therefore, it is conceivable to search for the optimum value by swarm intelligence search. Swarm intelligence search is a multi-point search algorithm that simulates organisms, etc., and is known to have high robustness that can search global optimum values with a small amount of calculation in a complex multimodal space. is there.

  However, the group intelligence search has a plurality of parameters, and the search performance changes depending on the values of the parameters. Therefore, it is necessary to tune the values of the parameters of the swarm intelligence search so that the search results satisfying the required performance are obtained. In the following, a set of parameters is referred to as a parameter set.

  In order to tune the parameter set for swarm intelligence search, if all the assumed parameter sets are solved in all the test cases that should be assumed, enormous time is required for tuning the parameter set. turn into.

  The present invention has been made based on this situation, and an object of the present invention is to provide a parameter adjustment device capable of shortening the calculation time for adjusting the value of the parameter set so as to satisfy the required performance. It is in.

  The above object is achieved by a combination of the features described in the independent claims, and the subclaims define further advantageous embodiments of the invention. Reference numerals in parentheses described in the claims indicate a correspondence relationship with specific means described in the embodiments described later as one aspect, and do not limit the technical scope of the present invention. .

  In order to achieve the above object, the present invention provides a parameter adjustment device for adjusting a value of a parameter set used in the first group intelligence search in order to search for a solution by the first group intelligence search, which satisfies the required performance. The ability value was determined by searching the solution by the first group intelligence search for the number of parameter sets corresponding to the number of low-density test parameter sets that is smaller than the total number of parameter sets determined based on the search range for searching the set. A first stage processing unit that applies the second group intelligence search based on the ability value and extracts the parameter set searched by the second group intelligence search as a parameter set candidate that is a candidate for the parameter set to be finally determined ( S4) and the second stage evaluation parameter set determined based on the parameter set candidates extracted by the first stage processing unit. A second stage processing unit (S5) for executing a high density test for evaluating the capability with the test cases set in advance for the number of high density test cases set as the number of test cases that can be confirmed to satisfy the second stage; A parameter determination unit (S5, S7) that determines a second-stage evaluation parameter set that satisfies the required performance as a result of the evaluation of the processing unit as a parameter set used for searching for a solution. A low density test for determining the capability value of the parameter set is executed with the number of test cases equal to the number of low density test cases which is smaller than the number of cases.

  According to the present invention, the first stage processing unit and the second stage processing unit are provided to evaluate the capability of the parameter set. Of these two processing units, it is only the second stage processing unit that evaluates the capability of the parameter set with the test cases for the number of high density test cases set as the number of test cases that can be confirmed to satisfy the required performance. is there.

  The first-stage processing unit executes a low-density test in which the ability value is determined by searching for a solution by the first group intelligence search for the parameter set in the number of low-density test cases less than the high-density test cases. To do. In addition, this low density test is only executed for the number of low density test parameter sets that is smaller than the total number of parameter sets determined based on the search range for searching for parameter sets that satisfy the required performance.

  That is, in the first stage extraction process, the number of test cases is reduced, and the number of parameter sets for determining the capability value is also reduced. Therefore, the time required for the first stage processing is shorter than the time required for executing the test cases for the number of high-density test cases with respect to the total number of parameter sets and determining the capability value of each parameter set. .

  However, in the first stage processing unit, the test cases are executed only for the number of low density test cases in which the number of test cases is reduced rather than the number of high density test cases. Therefore, the capability value determined by the first stage processing unit has an error with respect to the capability value when test cases corresponding to the number of high-density test cases are executed. Therefore, the final parameter set cannot be determined using the capability value determined by the first stage processing unit as it is. Therefore, a high density test is executed in the second stage processing unit.

  However, if a parameter set satisfying the required performance is not found as a result of executing the high-density test in the second stage processing, it is necessary to re-execute the first stage processing, and calculation until determining the parameter set satisfying the required performance There is a risk that the time cannot be shortened sufficiently.

  Therefore, it is necessary to increase the possibility that there is a parameter set that satisfies the required performance in the second-stage evaluation parameter set determined based on the parameter set candidates extracted by the first-stage processing unit. For this purpose, it is necessary to extract a large number of parameter set candidates in a range having a high ability value.

  Therefore, the first stage processing unit extracts the parameter set candidates by applying the second group intelligence search. In the first stage processing unit, since the capability value is determined by executing the low density test, there is an error in the capability value. However, a parameter whose capability value increases when a high density test is executed should have a high capability value even in a low density test. This is because the low density test is part of the high density test. Therefore, even if there is an error in the capability value determined by the low density test, the relationship between the capability value determined by executing the low density test and the parameter set is the capability value and parameter determined by executing the high density test. The relationship with the set is similar in trend.

  The swarm intelligence search is a method in which the vicinity of the optimum value is searched with priority. Therefore, the second group intelligence search is applied to extract parameter set candidates, and even if the low density test is executed and the ability value is determined, the ability value and the parameter set when the high density test is executed are determined. The relationship between the ability value of the tendency similar to the relationship and the parameter set candidate is obtained.

  In addition, a parameter set candidate is extracted by applying the second group intelligence search, that is, a method in which the vicinity of the optimum value is intensively searched. Therefore, even if the number of low-density parameter sets is set to a value sufficiently smaller than the total number of parameter sets, there is a high possibility that the second stage processing unit can extract parameter set candidates that can obtain an evaluation result that satisfies the required performance. If the number of low-density parameter sets is set to a sufficiently small value compared to the total number of parameter sets, the time required for the first stage processing can be reduced to the number of high-density test cases for the total number of parameter sets. The time required for determining the capability value of each parameter set can be much shorter.

  Then, the second stage processing unit performs a high density test on the second stage evaluation parameter set determined based on the parameter set candidate extracted in this way in the first stage processing unit, thereby improving the capability of the parameter set candidate. evaluate. Since the high density test is executed, it can be accurately determined whether the parameter set satisfies the required performance. In addition, since the parameter set for executing the high-density test is a second-stage evaluation parameter set determined based on the parameter set candidates extracted by the first-stage processing unit, the number of parameter sets for executing the high-density test is small. It is. Therefore, even if the high density test is executed, the calculation time is sufficiently short. As described above, in the present invention, the calculation time for searching for a parameter value satisfying the required performance can be shortened.

  In the invention according to claim 2, the second stage processing unit performs clustering to divide the parameter set candidates extracted by the first stage processing unit into a plurality of groups, and becomes a representative value of each of the plurality of groups. The parameter set is a second-stage evaluation parameter set.

  As described above, the relationship between the parameter set candidate extracted by the first stage processing unit and the capability value is similar in tendency to the relationship between the capability value determined by executing the high-density test and the parameter set.

  Therefore, when clustering for dividing the parameter set candidates extracted by the first stage processing unit into a plurality of groups is executed, the representative value of each group is highly likely to be the highest ability value in the group. Therefore, by setting the parameter set that is a representative value of each group as the second-stage evaluation parameter set, it is possible to reduce the possibility that there is no parameter set that satisfies the required performance as a result of executing the high-density test in the second-stage processing unit. As a result, there is less need to re-execute the first stage process, and the calculation time for searching for a parameter set that satisfies the required performance can be shortened.

  In addition, the second-stage evaluation parameter set is limited to parameter sets that are representative values of a plurality of groups. This also shortens the calculation time.

It is a figure which shows the structure of the electromagnetic wave arrival direction estimation system of embodiment. It is a figure which compares and shows a model waveform and a received waveform. It is a figure which illustrates the residual at the time of assuming that 2 waves have arrived. It is a figure which shows the parameter and search range with which particle group optimization is provided. It is a figure which shows the relationship between the fraction of an incoming wave, and the number of test cases. 3 is a flowchart showing PSO parameter setting processing determined by an azimuth angle determination unit 230 in FIG. 1. It is a flowchart which shows the process of S3 of FIG. 6 in detail. It is a flowchart which shows the process of S36 of FIG. 7 in detail. It is a figure explaining the PSO parameter extracted by the process of S3 of FIG. It is a flowchart which shows the process of S4 of FIG. 6 in detail. It is a figure explaining the clustering performed by S41 of FIG. It is a flowchart which shows the process of S46 of FIG. 10 in detail.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. The present embodiment is a radio wave arrival direction estimation system, and includes the wireless tag reader 1 and the wireless tag 300 shown in FIG. The wireless tag reader 1 has a function as a parameter adjustment device.

The wireless tag 300 transmits an unmodulated wave having a predetermined carrier frequency f ₀ set in advance. The wireless tag 300 is of an active type and may transmit radio waves continuously, but it is preferable to transmit radio waves intermittently from the viewpoint of battery life. The wireless tag 300 is carried by a person and has a size that can be easily accommodated in a pocket of clothes.

  The wireless tag reader 1 includes a reception unit 100 and a signal processing unit 200, and the reception unit 100 includes an antenna unit 110 and a low frequency signal generation unit 120.

[Description of Antenna Unit 110]
The antenna unit 110 includes an antenna 111, a turntable 112, and a drive unit 113. The antenna 111 is fixed to the outer peripheral edge of the turntable 112. There is no particular limitation on the shape and size of the antenna 111 as long as it can receive an unmodulated wave transmitted by the wireless tag 300 and can be fixed to a position other than the rotation center on the turntable 112.

  The turntable 112 is rotated by the drive unit 113. The shape of the rotating disk 112 is not limited to the disk shape, but is preferably not eccentric with respect to the driving unit 113. The turntable 112 is sized so that it can easily be set indoors. For example, a disk having a diameter of 10 cm. When the turntable 112 rotates, the antenna 111 fixed on it also rotates at the same time.

  The drive unit 113 includes a motor, and rotates the turntable 112 at a constant cycle. This fixed period is determined from the rotation speed of the antenna 111 determined from the Doppler shift to be secured and the rotation radius of the antenna 111.

[Description of Low Frequency Signal Generation Unit 120]
The low frequency signal generation unit 120 includes a band pass filter 121, a local oscillator 122, a mixer 123, a low pass filter 124, an A / D converter 125, a phase shifter 126, a mixer 127, a low pass filter 128, and an A / D converter 129. I have.

  The bandpass filter 121 uses a frequency range determined from a Doppler shift generated by rotation of the antenna 111 around the frequency of the radio wave transmitted by the wireless tag 300 as a pass frequency band. The band-pass filter 121 receives a reception signal received by the antenna 111 and removes noise from the reception signal. In addition, on the transmission path where the received signal is sent from the antenna 111 to the band pass filter 121, between one end of the rotation side transmission path rotating with the turntable 112 and one end of the non-rotation side transmission path facing the one end, The antenna pattern is transmitted oppositely by facing the antenna pattern. However, a slip ring may be used instead of wireless transmission.

The local oscillator 122 generates a local oscillation signal having the same frequency as the carrier frequency f ₀ transmitted by the wireless tag 300. The mixer 123 mixes the local oscillation signal and the signal output from the band-pass filter 121, and outputs a signal having a frequency that is the sum and difference of the frequency of the local oscillation signal and the frequency output from the band-pass filter 121. .

The low-pass filter 124 extracts a signal having a frequency that is the difference between the frequency of the local oscillation signal and the frequency output from the band-pass filter 121 from the signal output from the mixer 123. Since the frequency of the local oscillation signal is the same as the carrier frequency f ₀ transmitted by the wireless tag 300, the signal extracted by the low-pass filter 124 has a center frequency of 0 Hz. Hereinafter, the signal output from the low-pass filter 124 is referred to as an I component signal. The A / D converter 125 converts the I component signal that is an analog signal extracted by the low-pass filter 124 into a digital signal.

  The phase shifter 126 shifts the phase of the local oscillation signal by 90 °. The mixer 127 mixes the local oscillation signal whose phase is shifted by 90 ° by the phase shifter 126 and the signal output from the bandpass filter 121. The low pass filter 128 extracts a signal having a frequency that is the difference between the frequency of the local oscillation signal and the frequency output from the band pass filter 121 from the signal output from the mixer 127. However, the signal output from the low-pass filter 128 is 90 ° out of phase with the I component signal. A signal output from the low-pass filter 128 is hereinafter referred to as a Q component signal. This Q component signal also has a center frequency of 0 Hz. The A / D converter 129 converts the Q component signal, which is an analog signal extracted by the low-pass filter 128, into a digital signal. The I component signal and the Q component signal output from the A / D converters 125 and 129 correspond to the low frequency signal and the measurement signal in the claims.

[Description of Signal Processing Unit 200]
The signal processing unit 200 includes a signal acquisition unit 210, a storage unit 220, and an azimuth angle determination unit 230. The signal acquisition unit 210 acquires I component signals and Q component signals from the A / D converters 125 and 129, and stores the acquired signals in the storage unit 220.

  The wireless tag 300 transmits an unmodulated wave. However, the distance between the antenna 111 and the wireless tag 300 changes as the turntable 112 rotates. Therefore, the frequency of the radio wave received by the antenna 111 varies. Therefore, the frequency of the I component signal and the Q component signal acquired by the signal acquisition unit 210 also varies.

  The storage unit 220 stores an equation for estimating the radio wave arrival direction (hereinafter, radio wave arrival direction estimation equation). The radio wave arrival direction estimation formula is not limited as long as the radio wave arrival direction is provided as a variable. For example, the I component approximate model and the Q component approximate model described in Patent Document 1 can be used as the radio wave arrival direction estimation formula. Further, instead of the approximate model, an expression that strictly models the I component signal and the Q component signal may be used. Also, an equation obtained by modeling a received signal before being divided into IQ components may be used.

  FIG. 2 illustrates a model waveform and a received waveform. The reception waveform is a waveform that represents one of a reception signal, an I component signal, and a Q component signal. When the difference between the model waveform and the received waveform (hereinafter referred to as the residual) is minimized, that is, when the shape of the model waveform and the shape of the received waveform are the best match, It can be estimated that

  The azimuth angle determination unit 230 determines the arrival direction of the radio wave from the wireless tag 300 using the radio wave arrival direction estimation formula. Here, even if there is only one wireless tag 300, there are multiple directions of arrival of radio waves due to the influence of multipath. FIG. 3 illustrates the residual when it is assumed that two waves have arrived. As shown in FIG. 3, the residual space that represents the relationship between the arrival directions of the two radio waves and the residual is a very complex multimodal space. Thus, it is necessary to search for the minimum value instead of the minimum value in the residual space that is a multimodal space.

  Therefore, in the present embodiment, this minimum value, that is, the correct radio wave arrival direction is searched using group intelligence search. This is because the swarm intelligence search is known to have high robustness so that a global optimum can be searched with a small amount of calculation in a complex multimodal space.

  However, the search performance of the group intelligence search varies depending on the value of the parameter set. Therefore, it is necessary to tune the parameter set for swarm intelligence search. In order to tune the parameter set for swarm intelligence search, set a certain parameter set and perform direction estimation for a plurality of received waveform cases that are actually assumed. The search performance needs to be evaluated. The performance can be evaluated by the direction estimation accuracy and the calculation time.

  Here, the number of combinations of parameters set for the swarm intelligence search is generally enormous. In this embodiment, specifically, particle swarm optimization (hereinafter referred to as PSO) is used as the swarm intelligence search. As shown in FIG. 4, the PSO parameters (hereinafter referred to as PSO parameters) include the number of particles, group best term, self-best term, and inertia term. Therefore, the azimuth angle determination unit 230 tunes these PSO parameters. When these PSO parameters are changed between the minimum value and the maximum value shown in FIG. 4 with a necessary pitch width, there are tens of billions of combinations of PSO parameters. Millions of ways correspond to the total number of parameter sets in the claims.

  In addition, since it is necessary to assume a plurality of incoming waves by multipath, there are many conditions (hereinafter referred to as test cases) that need to be tested to evaluate the PSO parameter set. For example, as shown in FIG. 5, when testing the direction of arrival in increments of 10 degrees, if there are 1 wave, it is 36 directions, but if multiple waves are assumed, the number of test cases is generated by multiplying 36 by the wave number. To do.

  The time required for tuning the PSO parameter set is the product of “the number of combinations of PSO parameter sets”, “the number of test cases”, and “the operation time required for one test case”. Therefore, if all combinations of PSO parameter sets to be tested are evaluated for all test cases, an enormous amount of time is required.

  Therefore, the azimuth angle determination unit 230 of the present embodiment sets the PSO parameter set that can accurately estimate the arrival direction while reducing the “number of PSO parameter set combinations” and “number of test cases”. The PSO parameter setting process described in (1) is executed.

[Overview of PSO parameter setting process]
FIG. 6 shows the PSO parameter setting process executed in the present embodiment. First, an outline of the PSO parameter setting process will be described. The parameter setting process shown in FIG. 6 performs a two-stage extraction process of a first stage extraction process in step S4 (hereinafter, step is omitted) and a second stage extraction process in S5. The first stage extraction process corresponds to the first stage processing section of the claims, and the second stage extraction processing corresponds to the second stage processing section of the claims.

  In the first-stage extraction process, a small number of promising PSO parameters that are likely to satisfy the required performance are extracted while reducing the number of test cases (that is, at a low density). A small number of specific numbers to be extracted is, for example, thousands. Since there are tens of billions of combinations of PSO parameters in total, thousands are very few. These thousands are equivalent to the number of parameter sets for low-density testing.

  In order to extract a promising PSO parameter set, with a small number of test cases, the PSO parameter set is actually set in the PSO, the direction of arrival is estimated with the PSO, and the PSO capability is evaluated. Note that setting a PSO parameter set for a PSO, estimating the direction of arrival using the PSO, and evaluating the PSO capability is called a test, and the number of test cases to be tested is small (that is, the density is low). The case is called a low density test. In addition, high capability here means that in various test cases, the evaluation point determined by the arrival direction estimation accuracy and the calculation time is a high point.

  Even if the number of test cases is reduced and a small number of PSO parameter sets are extracted at random, the possibility that a promising PSO parameter set can be extracted is not high. Therefore, in the present embodiment, the second group intelligence search is used for extracting the PSO parameter set in the first stage extraction process. This is because swarm intelligence search is a technique that can search for an optimum value with a small amount of calculation even in a complex multimodal space.

  As described above, in the present embodiment, PSO is used for matching the model waveform. In order to distinguish this PSO from the second group intelligence search, it may be referred to as a first group intelligence search. The second group intelligence search here may be PSO as in the first group intelligence search, but may be differential evolution or cuckoo search. This is because differential evolution and cuckoo search are known to be methods that can easily extract a large number of local solutions.

  However, even if a promising PSO parameter set is extracted using the second group intelligence search, the first-stage extraction process has a small number of test cases, and thus there is an error in the ability value of the determined PSO parameter set. Finally, in order to evaluate whether the PSO satisfies the required performance, it is necessary to test without reducing the number of test cases. Tests that do not reduce the number of test cases are called high-density tests. In the second stage extraction process, this high density test is performed. In the second stage extraction process, the high-density test is performed only on representative values of each group obtained by clustering the promising PSO parameters extracted in the first stage extraction process. This is the end of the summary description. Hereinafter, the PSO parameter setting process will be described in detail with reference to FIGS.

[Detailed description of PSO parameter setting processing]
The process of FIG. 6 is executed as an initial setting after the wireless tag reader 1 is installed in the installation environment. In S1 of FIG. 6, the number P of particles for the second group intelligence search, the number of steps T _end , and the number of divisions K _cr for clustering are set. The number P of particles is a value set as appropriate based on experiments, and is, for example, 100 to 200. The number of steps T _end is a value representing how many times the second group intelligence search loop is executed. The clustering division number K _cr is the number of groups generated in the above-described clustering. The number of clustering divisions K _cr is, for example, about several tens.

These values are set by acquiring signals indicating these values from the outside of the signal processing unit 200. For example, the wireless tag reader 1 includes an input unit (not shown), and the user operates the input unit to input the number P of particles for the second group intelligence search, the number of steps T _end , and the number of divisions K _cr for clustering. To do. In this case, the number of particles P for the second group intelligence search, the number of steps T _end , and the number of clustering divisions K _cr are set in S1.

Further, it is connected to an external terminal in the wireless tag reader 1, the external terminals, the number of particles P in the second group intelligence search, the number of steps T _{end The,} the division number K _cr clustering may be input to the wireless tag reader 1.

  In S2, the search range and test case of the PSO parameter set are set. The search range of the PSO parameter set in S2 is from the minimum value to the maximum value of each parameter shown in FIG. Test cases are set for the low density test and the high density test, respectively.

  In each test case, at least the direction of arrival is set. The combinations of arrival directions include the number of combinations illustrated in FIG. Note that the number of waves to be assumed is set as appropriate according to the installation environment. In addition to the arrival direction, the phase and amplitude may be changed. However, the phase and amplitude may be fixed values.

  The number of test cases for the high-density test (that is, the number of high-density test cases) is the number of test cases shown in FIG. In addition, if the step and wave number in the direction of arrival are determined, a specific high-density test case is also determined. That is, a specific test case for performing a high density test is set in advance. On the other hand, the number of test cases for the low density test (that is, the number of low density test cases) is set based on the required calculation time.

  The number of low density test cases is, for example, 1/10 or less than the number of high density test cases. A specific test case for the low density test is a part of the test cases included in the test case for the high density test, and is set in advance. For example, if the direction of arrival is set in steps of 30 degrees, a specific test case for the low density test is included in the test case for the high density test. When the arrival direction is set to 30 degrees, assuming 3 waves, the number of test cases for 3 waves is 1728, which is 1/10 or less of the number of test cases for 3 waves shown in FIG.

  The setting in S2 is also set based on a user operation or connection of an external terminal, as in S1. Note that initial values are set in advance for the values set in S1 and S2, and only values that need to be changed may be changed. In continuing S3, a 1st step extraction process is performed.

[Description of the first stage extraction process]
Details of the first stage extraction process are shown in FIG. The first stage extraction process is a process for extracting promising PSO parameters using the second group intelligence search. The parameters for the second group intelligence search are set in advance. In FIG. 7, in S31, t is initialized, that is, 0. In S32, t is incremented by one. In S33, the particle number count CP is initialized, that is, 0. In S34, the particle number count CP is increased by one.

  In S35, a PSO parameter set is set in the PSO. As the PSO parameter set, a number of PSO parameter sets corresponding to the number of particles P set in S1 are randomly set as initial values.

  In S36, a low density test is executed. Details of this low density test are shown in FIG. In FIG. 8, in S361, i is initialized, that is, 0. In S362, i is incremented by one. In S363, the i-th test case i among the test cases set for the low-density test in S2 is solved with the PSO in which the PSO parameter set is set in S35. Solving test case i with PSO means that the arrival direction is estimated by matching the model waveform in the case of test case i using PSO. The received waveform for matching with the model waveform may be actually received, or may be created in advance by calculation such as simulation. PSO, which is a group intelligence search that actually solves a test case, corresponds to the first group intelligence search in the claims.

  In S364, the calculation result of S363 is evaluated according to the evaluation function. The calculation result here is the arrival direction estimated in S363 and the calculation time required for the processing in S363. The evaluation function is determined based on the required performance. The required performance is, for example, an arrival direction error of ± 3 degrees and a calculation time of 60 ms or less for all test cases. The evaluation function has a higher evaluation when the error in the arrival direction is smaller, and the evaluation is higher when the calculation time is earlier.

  In subsequent S365, the capability value for the currently used PSO parameter is updated. The capability value is a cumulative value of values (hereinafter referred to as evaluation values) obtained using the evaluation function in S364. Therefore, in S365, the evaluation value obtained in S364 of this time is added to the ability value so far.

  In S366, it is determined whether i has reached the number n of low density test cases. If this determination is NO, the process returns to S362. Therefore, in the process of FIG. 8, the PSO capability value when the PSO parameter set in S35 of FIG. 7 is set is determined by testing the number of low density test cases. If the determination in S366 is YES, the process in FIG. 8 is terminated, and the process proceeds to S37 in FIG.

In S37, it is determined whether or not the particle number count CP is equal to the particle number count end value CP _end . The particle count end value CP _end is the same value as the particle count P set in S1. If this determination is NO, the process returns to S34.

  In S34, as described above, the particle count value is increased by 1. Therefore, in the subsequent S35, the PSO parameter set corresponding to the increased particle number count CP is set to PSO. In S36, a low density test is executed with the PSO. Therefore, by repeating S34 to S37, the capability values of the PSO parameter sets corresponding to the number P of particles are determined by the low density test.

  If judgment of S37 becomes YES, it will progress to S38. In S38, the PSO parameter set of each particle is updated. In this update, for each particle, the best ability value so far is compared with the ability value obtained this time to update the best ability value. Then, the PSO parameter set of each particle is updated on the basis of the PSO parameter set that is the best updated capability value. Therefore, the vicinity of the best ability value is intensively searched. Thus, the swarm intelligence search focuses on the vicinity of the best value. Therefore, even if the number of combinations of PSO parameter sets to be extracted is small, a promising (that is, high capability value) PSO parameter set can be extracted.

In S39, t is judged whether equal to or greater than the number of steps _{T end} set in S1. If this determination is NO, the process returns to S32. Therefore, S32 to S38 is repeated by the number of steps _{T end The.} When returning to S32, in S35 to be executed next, the PSO parameter set updated in the immediately preceding S38 is set.

In addition, S34 to S37 are repeated by the number of particles P while S32 to S38 are executed once. When S34 to S37 are executed once, the capability value of the PSO parameter set set in S35 is reduced to the low density test. Determined by Therefore, by executing the first stage extraction process shown in FIG. 7, the capability values of the PSO parameter sets corresponding to the number of particles P × the number of steps T _end are determined. The number of particles P × the number of steps T _end corresponds to the number of parameter sets for low density test. Specifically, this number is several thousand sets as described above.

  In FIG. 9, the PSO parameters extracted by the first stage extraction process are indicated by black circles. Further, the broken line in FIG. 9 indicates the capability of the PSO parameter determined by executing all the test cases for all combinations of the PSO parameters.

  As can be seen from the fact that the black circle is deviated from the broken line, the ability of each PSO parameter determined in the first stage extraction process has an error. The reason for the error is that the ability was determined by running a low density test with a small number of test cases.

  If the PSO parameter to be finally adopted is determined based on the ability having an error, there is a possibility that a PSO parameter set that satisfies the required performance cannot be set. Therefore, when the first stage extraction process is completed, the second stage extraction process is executed in S4 of FIG.

[Description of the second stage extraction process]
Details of the second stage extraction process are shown in FIG. In FIG. 10, in S41, the PSO parameter sets extracted in the first stage extraction process are clustered. The PSO parameter set extracted in the first stage extraction process is a parameter set candidate.

  The k-means method is used for clustering. The number of divisions (that is, the number of clusters) is the number set in S1. The PSO parameter set includes four types of parameters as shown in FIG. The PSO parameters to be divided here may be all four types, but may be only the three types of parameters excluding the number of particles. This is because it can be estimated that the larger the number of particles, the better the result, and therefore it is not necessary to extract the median value in the next S42.

  FIG. 11 conceptually shows the clustering result in S41. Since the actual clustering results are obtained by clustering three or four types of PSO parameters, there are three or four parameter axes. In FIG. 11, the range enclosed by a square is one group.

  In S42, for each group divided in S41, a PSO parameter set serving as the median value of each group is extracted. The PSO parameter set extracted here is the second stage evaluation parameter set in the claims.

  In clustering by the k-means method, clustering is performed while calculating the distance from the center value. This central value is used as the central value here. The median is a representative value of each group. In FIG. 11, the median value of each group is indicated by a solid line within a square indicating each group. When only three types of parameters excluding the number of particles are targeted for clustering, the number of particles is set to a predetermined constant value.

  In S43, k is initialized, that is, 0 is set. In S44, k is increased by 1. In S45, the PSO parameter for the kth group among the PSO parameters extracted in S42 is set in the PSO. In S46, a high density test is executed. Details of this high density test are shown in FIG.

  In FIG. 12, i is initialized, that is, set to 0 in S461. In S462, i is incremented by one. In S463, the i-th test case i among the test cases set for the high-density test in S2 is solved with the PSO in which the PSO parameter set is set in S45.

  In S464, the calculation result of S463 is evaluated according to the evaluation function. The calculation result here is the arrival direction estimated in S463 and the calculation time required for the processing in S463. The evaluation function is the same as in the low density test.

  In subsequent S465, the capability value for the PSO parameter set being used is updated. Specifically, the evaluation value obtained in S464 this time is added to the ability value so far. In S466, it is determined whether i has reached the number m of high-density test cases. If this determination is NO, the process returns to S462. Therefore, in the process of FIG. 12, the PSO capability value when the PSO parameter set set in S45 of FIG. 10 is set is determined by testing the number of high-density test cases. If the determination in S466 is YES, the process in FIG. 12 is terminated, and the process proceeds to S47 in FIG.

In S47, it is determined whether or not k is equal to or greater than the number of divisions K _cr set in S1. If this determination is NO, the process returns to S44. Therefore, S44～S47 is repeated by the number of divisions _{K cr.} Thereby, the ability value of the PSO parameter set that is the median value of each group extracted in S42 is obtained. If the determination in S47 is YES, the process proceeds to S5 in FIG.

In S5, it is determined whether there is a PSO parameter that satisfies the required performance in the PSO parameter set that is the median value of each group extracted in the second stage extraction process. The required performance is set in advance. For example, for all test cases used in the high-density test, an arrival direction error of ± 3 degrees and a calculation time of 60 ms or less are satisfied. If judgment of S5 is NO, it will progress to S6. In S6, the setting is changed. How to change the setting may be determined by the user as appropriate, but a change is made to increase one or both of the number of particles P and the number of steps _Tend . After changing the setting, S3 and subsequent steps are executed with the changed setting value.

  On the other hand, if judgment of S5 is YES, it will progress to S7. In S7, the PSO parameter set having the highest capability value among the PSO parameter sets extracted in S42 of the second stage extraction process is determined as the PSO parameter set used for estimating the radio wave arrival direction. Note that S5 and S7 correspond to a parameter determination unit in claims.

  The azimuth angle determination unit 230 sets the determined PSO parameter set to PSO. When the received waveform is acquired, the model waveform that most closely matches the received waveform is searched by the PSO, and the arrival direction (that is, the azimuth angle) of the radio wave is sequentially estimated.

[Summary of Embodiment]
As described above, in the present embodiment, the first stage process and the second stage extraction process are performed in order to evaluate the capability of the PSO parameter set. Of these two extraction processes, only the second stage extraction process evaluates the PSO parameter set capability by performing a high density test.

  In the first stage process, a low density test is executed. The low-density test is a test that determines the capability value of the parameter set with as many test cases as the number of low-density test cases than the high-density test case, and the number of parameter sets to be tested is more than the total number of parameter sets. Overwhelmingly few. Therefore, the time required for the first stage extraction process can be much shorter than the time required to execute the high density test for the total number of parameter sets.

  However, in the first stage extraction process, since the capability value is determined by the low density test, the determined capability value has an error. Therefore, a final PSO parameter set cannot be determined using the capability value determined in the first stage extraction process as it is.

  Therefore, the high-density test is executed in the second-stage extraction process. If the PSO parameter set satisfying the required performance is not found as a result of executing the high-density test in the second-stage extraction process, the first-stage extraction process is performed. It becomes necessary to re-execute. Therefore, it is necessary to increase the possibility that there is a parameter set that satisfies the required performance in the PSO parameter set that executes the high-density test in the second stage extraction process. Therefore, it is necessary to extract a large number of parameter sets having a high capability value in the first stage extraction process.

  Therefore, in the first stage extraction process, the PSO parameter set is extracted by applying the second group intelligence search. In the first stage extraction process, since the capability value is determined by executing the low density test, the capability value has an error. However, since the low density test is a part of the high density test, a parameter whose capability value becomes high when the high density test is executed should have a high capability value even in the low density test.

  Therefore, even if there is an error in the capability value determined by the low density test, the relationship between the capability value determined by executing the low density test and the PSO parameter set is the same as the capability value determined by executing the high density test. The relationship with the parameter set is similar in trend. This can be seen from the comparison between the broken line and the black circle in FIG.

  The swarm intelligence search is a method in which the vicinity of the optimum value is intensively searched. Therefore, by extracting the PSO parameter set by applying the second group intelligence search, even if the ability value is determined by executing the low density test, the ability when the ability value is determined by executing the high density test It is possible to obtain the relationship between the capability value of the tendency similar to the relationship between the value and the PSO parameter set and the PSO parameter set.

  In addition, the PSO parameter set is extracted by applying the second group intelligence search. Thereby, even if the number of PSO parameter sets extracted in the first stage extraction process is set to a value sufficiently smaller than the total number of parameter sets, a PSO parameter set that can obtain an evaluation result that satisfies the required performance in the second stage extraction process can be obtained. Is likely to be extracted.

  If the number of PSO parameter sets to be extracted in the first stage extraction process is set to a value sufficiently smaller than the total number of parameter sets, the time required for the first stage extraction process can be increased with respect to the total number of parameter sets. The time can be overwhelmingly shorter than the time required to execute the test and determine the capability value of each PSO parameter set.

  In the second stage extraction process, a high density test is performed on the PSO parameter set determined from the PSO parameter set extracted in this way in the first stage extraction process to evaluate the capability of the PSO parameter set. Since the high density test is executed, it can be accurately determined whether the parameter set satisfies the required performance. In addition, the parameter set for executing the high density test is only several tens of parameter sets in the present embodiment. Therefore, even if the high density test is executed, the calculation time is sufficiently short. As described above, in the present embodiment, the calculation time for searching for the value of the PSO parameter set that satisfies the required performance can be shortened.

  Further, in the present embodiment, clustering for dividing the PSO parameter set extracted in the first stage extraction process into a plurality of groups is executed, and the high-density test is executed only on the PSO parameter set that is a representative value of each group.

  Since the relationship between the PSO parameter set extracted in the first stage extraction process and the capability value is similar in tendency to the relationship between the capability value determined by executing the high-density test and the parameter set, the representative value of each group is There is a high possibility that it will be the highest ability value in the group. Therefore, in the present embodiment, the capability value is determined by executing the high density test only for the PSO parameter set that is a representative value of each group.

  Thereby, the calculation time required for the second stage extraction process can be shortened. In addition, since it is possible to reduce the possibility that there is no parameter set that satisfies the required performance as a result of the second stage extraction process, the need for re-execution of the first stage extraction process is reduced. From these facts, the calculation time for searching for a parameter set that satisfies the required performance can be further shortened.

  As mentioned above, although embodiment of this invention was described, this invention is not limited to the above-mentioned embodiment, The following modification is also contained in the technical scope of this invention, Furthermore, the summary other than the following is also included. Various modifications can be made without departing from the scope.

<Modification 1>
For example, in the above-described embodiment, the present invention is applied to the radio wave arrival direction estimation system. However, the present invention is not limited to the radio wave arrival direction estimation system, and can be widely applied to techniques for obtaining a solution by group intelligence search.

<Modification 2>
In the above-described embodiment, clustering is performed by the k-means method, but other clustering methods may be used.

<Modification 3>
In the above-described embodiment, the PSO is used as the first group intelligence search. However, the first group intelligence search may be a group intelligence search other than the PSO.

<Modification 4>
In the above-described embodiment, the PSO parameter sets extracted in the first stage extraction process are clustered, and the high density test is executed on the PSO parameter sets that are representative values of each group. However, the PSO parameter set extracted in the first stage extraction process may be used as the PSO parameter set for executing the high density test as it is.

DESCRIPTION OF SYMBOLS 1: RFID tag reader 100: Reception part 110: Antenna part 111: Antenna 112: Turntable 113: Drive part 120: Low frequency signal generation part 121: Band pass filter 122: Local oscillator 123: Mixer 124: Low pass filter 125: A / D converter 126: Phase shifter 127: Mixer 128: Low pass filter 129: A / D converter 200: Signal processing unit 210: Signal acquisition unit 220: Storage unit 230: Azimuth angle determination unit 300: Radio tag

Claims (2)

A parameter adjustment device for adjusting a value of a parameter set used in the first group intelligence search in order to search for a solution by the first group intelligence search,
The solution is searched by the first group intelligence search for the number of parameter sets corresponding to the number of low-density test parameter sets smaller than the total number of parameter sets determined based on the search range for searching for the parameter sets satisfying the required performance. Determining a capability value, applying a second group intelligence search based on the determined capability value, and determining the parameter set searched by the second group intelligence search as a parameter set candidate to be finally determined A first stage processing unit (S4) for extracting as a parameter set candidate
For the second-stage evaluation parameter set determined based on the parameter set candidates extracted by the first-stage processing unit, the number of high-density test cases set as the number of test cases that can be confirmed to satisfy the required performance A second stage processing unit (S5) for executing a high density test for evaluating the capability in a preset test case;
A parameter determination unit (S5, S7) for determining, as the parameter set used for the solution search, the second-stage evaluation parameter set in which the evaluation result of the second-stage processing unit satisfies the required performance;
The first stage processing unit adjusts parameters for executing a low density test for determining a capability value of the parameter set in test cases corresponding to the number of low density test cases, which is the number of test cases smaller than the number of high density test cases. apparatus. In claim 1,
The second stage processing unit performs clustering to divide the parameter set candidates extracted by the first stage processing unit into a plurality of groups, and sets the parameter sets that are representative values of the plurality of groups, A parameter adjustment device for the second stage evaluation parameter set.