JP7093965B2 - Model selection device and model selection method - Google Patents

Description

The present invention relates to a model selection device and a model selection method.

In machine learning that supports modern artificial intelligence (AI), a technology called "reinforcement learning" is attracting attention along with a technology called "deep learning" that demonstrates excellent abilities in image recognition and the like. A central issue in reinforcement learning is, for example, how to maximize profits in a casino with a large number of slot machines. In order to maximize profits, it is necessary to carry out trial and error (trial and error) to find a "hit stand". However, excessive trial strikes can lead to losses, and the hitting platform can change from moment to moment. On the other hand, if you stop the trial shot early, you may miss the important point. Thus, there is a difficult dilemma between "searching" and "deciding."

The above problem is known as the "multi-armed bandit problem", and is used to solve various problems such as frequency allocation in wireless communication, computational resource allocation in data centers, robot control, and web advertising. Various studies are being conducted because it is expected to be applied.

In research on a method for solving a multi-armed bandit problem, it is assumed that a machine with a high hit probability is selected from a large number of slot machines, and principle verification and evaluation performance are performed. An ultra-high-speed method using chaos has been proposed.

M. Naruse, Y. Terashima, A. Uchida, and S. -J. Kim: Ultrafast photonic reinforcement learning based on laser chaos, SCIENTIFIC REPORTS, Vol. 7, Article No. 8772, August 2017.

However, the disclosure techniques so far have been limited to studies targeting virtual slot machines, and are limited to targeting specific events. For this reason, the challenge is to develop the solution of problems in various realistic events.

Therefore, the present invention has been devised in view of the above-mentioned problems, and an object thereof is a model selection device and a model selection method that can be used for solving problems in various realistic events. Is to provide.

The model selection device according to the first invention is based on an acquisition unit that acquires an irregular signal in which at least two types of characteristic signal strengths change irregularly, and a first signal strength of the irregular signal at the first timing. Acquires a plurality of estimation units that generate a prediction series that estimates the signal strength of the irregular signal at the second timing after the first timing, and the second signal strength of the irregular signal at the second timing. Then, based on the evaluation unit that calculates the prediction error of each of the plurality of prediction series with respect to the second signal strength and the plurality of prediction errors, it is suitable for estimating the second signal strength among the plurality of estimation units. It is provided with a search unit for searching for a superior estimation unit, and each of the plurality of estimation units is optimized for estimating the characteristic signal strength of one type of the irregular signal. It is a feature.

In the model selection device according to the second invention, in the first invention, the search unit has a threshold value adjusting unit that adjusts a threshold value for selecting the superior estimation unit based on a plurality of prediction errors, and an irregular signal. The superiority estimation is based on a signal transmitting unit that emits a selection irregular signal whose intensity changes, a comparison unit that compares the signal intensity of the selection irregular signal with the threshold value, and a comparison result by the comparison unit. It is characterized by having a selection unit for selecting a unit.

In the first invention or the second invention, the model selection device according to the third invention further includes a conversion unit that converts the first signal intensity into an optical signal, and the plurality of estimation units are output from the conversion unit. It is characterized in that the prediction sequence is generated based on an optical signal.

In the third aspect of the invention, the model selection device according to the fourth aspect of the present invention obtains different weight information in advance from each of the plurality of estimation units, and calculates the weight information in the virtual node state acquired from the output of the optical signal. The prediction series is generated, and the weight information is characterized in that it is constructed to estimate the characteristic signal strength of one of the irregular signals.

The model selection device according to the fifth aspect of the invention is characterized in that, in any one of the first to fourth inventions, the number of the plurality of estimation units is equal to or greater than the number of the characteristic signal strength types.

The model selection method according to the sixth invention is based on an acquisition step of acquiring an irregular signal in which at least two kinds of characteristic signal intensities change irregularly, and a first signal intensity of the irregular signal at the first timing. , The estimation step of generating a plurality of prediction sequences for estimating the signal strength of the irregular signal at the second timing after the first timing, and the second signal strength of the irregular signal at the second timing are acquired. , An evaluation step for calculating the prediction error of each of the plurality of prediction series with respect to the second signal strength, and the second signal strength of the plurality of estimation units that generated the prediction series based on the plurality of prediction errors. Each of the plurality of prediction sequences is generated based on one of the irregular signals suitable for estimating the characteristic signal strength, including a search step for searching for a superior estimation unit suitable for estimation. It is characterized by being a product.

According to the present invention having the above-described configuration, the search unit searches for a superior estimation unit suitable for estimating the second signal strength based on a plurality of prediction errors. Further, each of the plurality of estimation units is optimized for estimating the characteristic signal strength of one type of the irregular signal. Therefore, regardless of the type and timing of the irregular signal, it is possible to search for the estimation unit (dominant estimation unit) most suitable for the prediction performance. This makes it possible to use it to solve problems in various realistic events.

Further, in the present invention having the above-mentioned configuration, an approach of selecting the most suitable estimation unit (model) for the actual situation is realized by using a plurality of estimation units formed so as to maximize the performance for a specific aspect. can do. This does not require enormous learning or acquisition of generalization ability, and it is possible to achieve both short latency and high throughput.

Further, according to the present invention having the above-described configuration, the search unit may include a threshold value adjusting unit, a signal transmitting unit, a comparison unit, and a selection unit. Therefore, when searching for the superiority estimation unit, reinforcement learning using optical chaos can be used. This makes it possible to shorten the search time of the superiority estimation unit.

In particular, the threshold adjustment unit may adjust the threshold value for selecting the superior estimation unit based on a plurality of prediction errors. Therefore, even when there is a large variation in the prediction error, it is possible to carry out the selection following the selection tendency up to the immediately preceding, and it is possible to reduce the selection error due to the variation in the prediction error. This makes it possible to suppress a decrease in the accuracy of selecting the superior estimation unit for an irregular signal that changes irregularly.

Further, according to the present invention having the above-described configuration, the conversion unit may convert the first signal intensity into an optical signal. Therefore, the time for generating the prediction series can be significantly shortened as compared with the case where the first signal intensity is converted into an object other than the optical signal. This makes it possible to search for the superior estimation unit most suitable for the prediction performance in a short time.

Further, according to the present invention having the above-described configuration, the plurality of estimation units generate a prediction series based on an optical signal. Therefore, it is possible to easily realize a combination of a conversion unit that can generate a prediction series in a short time and a search unit that uses reinforcement learning. This makes it possible to significantly shorten the search time of the superiority estimation unit.

Further, according to the present invention having the above-described configuration, the weight information may be constructed to estimate the characteristic signal strength of one of the irregular signals. Therefore, it can be easily constructed as compared with the case of constructing one weight information for estimating a plurality of characteristic signal strengths. This makes it possible to simplify advance preparation.

Further, according to the present invention having the above-described configuration, the number of estimation units may be greater than or equal to the number of characteristic signal strength types. Therefore, it is possible to search for all the characteristic signal intensities included in the irregular signal under the condition that the estimation unit most suitable for the prediction performance is always included. This makes it possible to improve the search accuracy.

According to the present invention having the above-described configuration, the search step searches for a superior estimation unit suitable for estimating the second signal strength based on a plurality of prediction errors. Further, each of the plurality of prediction series is generated based on the characteristic signal strength of one type of the irregular signal suitable for estimation. Therefore, regardless of the type and timing of the irregular signal, it is possible to search for the estimation unit (dominant estimation unit) most suitable for the prediction performance. This makes it possible to use it to solve problems in various realistic events.

It is a schematic diagram which shows an example of the whole structure of the model selection apparatus to which this invention is applied. It is a schematic diagram which shows an example of a search part. It is a schematic diagram which shows another example of a search part. It is a block block diagram of the signal transmission part in the case of generating a return semiconductor laser light. It is a figure which shows the process of detecting the signal strength in a decision-making part. It is a schematic diagram which shows the other example of the whole structure of the model selection apparatus to which this invention is applied. It is a flowchart which shows an example of the model selection method to which this invention is applied. (A) is a graph showing the irregular signal used in the example, and (b) and (c) are graphs showing the prediction series generated in the example. (A) and (b) are graphs showing the prediction error calculated in the examples. (A) is a graph showing the transition of the threshold value in the embodiment, and (b) is a graph showing the ratio of correctly selecting the selection unit suitable for estimation as the superior estimation unit.

Hereinafter, the model selection device to which the present invention is applied will be described in detail with reference to the drawings.

(Configuration of model selection device 1)
FIG. 1 is a schematic diagram showing an example of the overall configuration of the model selection device 1 to which the present invention is applied. The model selection device 1 can search for the estimation unit 12 (dominant estimation unit) most suitable for the prediction performance for the irregular signal TS, and can be used to solve problems in various realistic events. Become.

The model selection device 1 includes an acquisition unit 11, an estimation unit 12, an evaluation unit 13, and a search unit 14, and may include a conversion unit 15, for example, as shown in FIG.

<Acquisition unit 11>
The acquisition unit 11 acquires the irregular signal TS. The acquisition unit 11 may, for example, apply arbitrary mask processing to the irregular signal TS. The acquisition unit 11 outputs, for example, the signal strength of the irregular signal TS at an arbitrary timing. The acquisition unit 11 outputs, for example, the first signal strength TS1 at the first timing and outputs the second signal strength TS2 at the second timing after the first timing. The interval between the first timing and the second timing is arbitrary.

The irregular signal TS is a signal in which at least two types of characteristic signal intensities change irregularly. As the characteristic signal strength included in the irregular signal TS, for example, a signal strength formed based on a Lorentz type time series, a wrestler type time series, or the like may be used.

<Estimating unit 12>
The estimation unit 12 generates a prediction sequence PW that estimates the signal strength of the irregular signal TS, and a plurality of estimation units 12 are provided. The estimation unit 12 generates a prediction sequence PW that estimates the signal strength of the irregular signal TS (second signal strength TS2) at the second timing based on, for example, the first signal strength TS1. Each of the plurality of estimation units 12 is optimized for estimating the characteristic signal strength of one type of the irregular signal TS. Therefore, each estimation unit 12 (for example, the first estimation unit 12a and the second estimation unit 12b) has different prediction series PWs (for example, the first prediction series PW1 and the second prediction series PW1) based on the same first signal strength TS1. Generate PW2). The number of estimation units 12 is, for example, greater than or equal to the number of characteristic signal strength types.

<Evaluation unit 13>
The evaluation unit 13 acquires the second signal strength TS2 of the irregular signal TS at the second timing. The evaluation unit 13 calculates each prediction error PE of the plurality of prediction series PW with respect to the second signal strength TS2. The prediction error PE is calculated, for example, by the difference between the second signal strength TS2 and the prediction series PW. That is, the prediction error PE is calculated as an index indicating the prediction accuracy of each estimation unit 12.

<Search unit 14>
The search unit 14 searches for an estimation unit 12 (dominant estimation unit) suitable for estimating the second signal strength TS2 among the plurality of estimation units 12 based on the plurality of prediction error PEs. For example, the search unit 14 may search for the estimation unit 12 that generated the prediction series PW having the smallest absolute value of the prediction error PE as the superior estimation unit.

As shown in FIG. 2, for example, the search unit 14 includes a threshold value adjustment unit 14a, a signal transmission unit 14b, a comparison unit 14c, and a selection unit 14d. The search unit 14 may search the estimation unit 12 via the threshold value adjustment unit 14a, the signal transmission unit 14b, and the comparison unit 14c.

<Threshold adjustment unit 14a>
The threshold value adjusting unit 14a adjusts the threshold value TH for selecting the estimation unit 12 (dominant estimation unit) suitable for estimation based on the plurality of prediction error PEs. In order to adjust the threshold value TH before adjustment, the threshold value adjusting unit 14a forms the threshold value TH in a state where the content of the newly generated prediction error PE is added after following the selection tendency until immediately before. Can be done. The threshold value adjusting unit 14a adjusts the threshold value TH so that, for example, the estimation unit 12 that generated the prediction error PE having the smallest absolute value is preferentially selected.

<Signal transmitter 14b>
The signal transmitting unit 14b transmits an irregular selection signal whose signal strength changes irregularly. The irregular selection signal transmitted from the signal transmission unit 14b is supplied to the decision-making unit 4 including the comparison unit 14c and the selection unit 14d.

The selection irregular signal transmitted by the signal transmitting unit 14b is not limited to the case where the signal strength simply changes randomly, and is not limited to the case where the signal strength changes chaotically. That is, the irregular signal for selection may be composed of a signal having complete irregularity having no correlation with the signal strength at the signal transmission time immediately before the signal transmission time. It may be composed of a signal having a part of irregularity having a slight correlation with the signal strength at the time of signal transmission immediately before the time of signal transmission.

It is not essential for the signal transmitting unit 14b to transmit an electrical or physical signal, and the data string of the previously acquired irregular signal for selection may be supplied to the decision-making unit 4. For example, the selection irregular signal transmitted from the signal transmitting unit 14b is based on the return semiconductor laser light expressed by returning the light emitted from the semiconductor laser into the laser resonator via the reflecting unit. It may be generated.

FIG. 4 shows a block configuration of the signal transmitting unit 14b in the case of generating the return semiconductor laser light. The signal transmitting unit 14b includes a semiconductor laser 91, a coupler 92, an attenuator 93, a reflecting unit 94, an optical isolator 95, an attenuator 96, a light receiving unit 97, and an oscilloscope 98.

The semiconductor laser 91 oscillates a laser beam based on a supplied electric signal. The coupler 92 plays a role of branching the laser light from the semiconductor laser 91 and the return light from the attenuator 93 and the reflecting unit 94. The attenuator 93 attenuates the return light from the reflecting unit 94. The reflecting unit 94 is composed of a mirror or the like that reflects the laser beam from the semiconductor laser 91. The optical isolator 95 plays a role of transmitting only the component of light traveling in the forward direction and blocking the component of light in the opposite direction with respect to the laser light emitted from the coupler 92. The attenuator 96 attenuates the laser beam from the optical isolator 95. The light receiving unit 97 is composed of a photodiode or the like, and photoelectrically converts the laser light from the optical isolator 95 into an electric signal. The oscilloscope 98 uses the electrical signal photoelectrically converted by the light receiving unit 97 as a base on the time axis, and waveforms the signal intensity thereof.

According to the signal transmitting unit 14b having such a configuration, the laser light oscillated by the semiconductor laser 91 first reaches the reflecting unit 94 via the coupler 92 and the attenuator 93. This laser light is reflected by the reflecting unit 94, the amount of return light is adjusted by the attenuator 93, and then the laser light is returned to the semiconductor laser 91 as return light via the coupler 92 again. The return light returned to the resonator of the semiconductor laser 91 significantly increases the noise level of the light and destabilizes it, and as a result, the return semiconductor laser light is emitted again from there. This return semiconductor laser light has the above-mentioned irregular signal because the coherence has collapsed, so to speak, in a laser chaotic state. Such return semiconductor laser light is attenuated in the attenuator 96 via the coupler 92 and the optical isolator 95, photoelectrically converted in the light receiving unit 97, and further waveformized in the oscilloscope 98. The waveformized signal in this oscilloscope 98 is an irregular signal. The signal transmission unit 14b supplies such an irregular signal as a selection irregular signal to the decision-making unit 4.

<Decision-making unit 4, comparison unit 14c, selection unit 14d>
The decision-making unit 4 detects the signal strength of the irregular selection signal transmitted from the signal transmission unit 14b. The decision-making unit 4 includes a comparison unit 14c and a selection unit 14d. The comparison unit 14c compares the signal strength of the irregular selection signal with the threshold value TH. The selection unit 14d selects the estimation unit 12 (dominant estimation unit) based on the comparison result by the comparison unit 14c. For example, when there are two estimation units 12 (FIG. 2), the threshold value TH used by the comparison unit 14c is one (TH ₁ ), and the selection unit 14d selects one from the two estimation units 12a and 12b. For example, when there are four estimation units 12 (FIG. 3), the threshold values TH used by the comparison unit 14c are three (TH ₁ , TH ₂ , 0, TH 2, ₁ ), and the selection unit 14d has four estimation units 12a, Select one from 12b, 12c, and 12d.

FIG. 5 shows a process of detecting the signal strength in the decision-making unit 4. Note that FIG. 5 shows an example of selecting the superior estimation unit from the four estimation units 12a, 12b, 12c, and 12d.

As shown in FIG. 5, when the selection irregular signal is time on the horizontal axis and the vertical axis is the signal strength, the comparison unit 14c detects the signal strength of the selection irregular signal at each of two or more observation time points. do. For example, when this observation time point is set to two time points, t ₁ and t ₂ , the signal strength P ₁ of the irregular selection signal at the observation time point t ₁ is detected, and the selection time point t ₂ is not selected. The signal strength P ₂ of the regular signal is detected.

Next, the comparison unit 14c compares the signal strength detected at each observation time point with the threshold value TH set for each observation time point. Here, it is assumed that the threshold value TH ₁ is preset at the observation time point t ₁ , and the threshold values TH 2, ₀ and TH 2, ₁ are preset at the observation time point t ₂ . First, the signal strength P ₁ of the irregular signal for selection at the observation time t ₁ and the threshold value TH ₁ are compared. As a result, when the signal strength P ₁ is smaller than the threshold value TH ₁ , the process proceeds in the arrow direction of the threshold value TH 2, ₀ , and when the signal strength P ₁ is equal to or higher than the threshold value TH ₁ , the arrow direction of the threshold value TH ₂ , 1. Proceed to. When the threshold value TH _2,0 is advanced in the direction of the arrow, the signal strength P ₂ of the selection irregular signal at the observation time point t ₂ is compared with the threshold value TH 2, ₀ . The selection unit 14d selects the estimation unit 12 according to the magnitude relationship between the signal strength P ₂ and the threshold value TH _2,0 . Similarly, when the threshold value TH ₂ , 1 is advanced in the direction of the arrow, the signal strength P ₂ of the selection irregular signal at the observation time point t ₂ is compared with the threshold value TH 2, ₁ . The selection unit 14d selects the estimation unit 12 according to the magnitude relationship between the signal strength P ₂ and the threshold values TH 2, ₁ .

Here, the estimation units 12a, 12b, 12c, and _12d described above are pre-assigned with threshold values TH ₁ , TH 2, ₀ , and TH 2, 1, respectively. For example, the estimation unit 12a is assigned below the threshold TH _2,0 , the estimation unit 12b is assigned above the threshold TH _2,0 , the estimation unit 12c is assigned below the threshold TH _2,1 , and the estimation unit 12d is assigned. , It is assumed that they are assigned at the threshold value TH _2,0 or higher. As a result of comparison between the signal strength P ₂ of the irregular signal for selection and the threshold value TH, the estimation unit 12 to be selected corresponds to the estimation unit 12 assigned in advance across each threshold value TH.

In the case of the above-mentioned example, the selection unit 14d selects the estimation unit 12a when the signal intensity P ₂ is smaller than the threshold value TH _2.0 . When the signal intensity P ₂ is equal to or higher than the threshold value _{TH 2} , the selection unit 14d selects the estimation unit 12b. When the signal intensity P ₂ is smaller than the threshold value TH 2, ₁ , the selection unit 14d selects the estimation unit 12c. When the signal intensity P ₂ is the threshold value TH 2, ₁ or more, the selection unit 14d selects the estimation unit 12d.

That is, the case where the estimation unit 12a is selected as the dominant selection unit is a case where the signal strength P ₁ is less than the threshold value TH ₁ and the signal strength P ₂ is less than the threshold value TH _2.0 . The case where the estimation unit 12b is selected as the dominant estimation unit is a case where the signal strength P ₁ is less than the threshold value TH ₁ and the signal strength P ₂ is the threshold value TH _2.0 or more. The case where the estimation unit 12c is selected as the dominant estimation unit is a case where the signal strength P ₁ is equal to or higher than the threshold value TH ₁ and the signal strength P ₂ is lower than the threshold value TH _{2 1} . The case where the estimation unit 12d is selected as the dominant estimation unit is a case where the signal intensity P ₁ is the threshold value TH ₁ or more and the signal intensity P ₂ is the threshold value TH 2, ₁ or more. When the estimation units 12a and 12b are actually selected, the threshold values TH ₁ and the threshold values TH 2, ₀ are used, and the threshold values TH 2, ₁ are not particularly used. Similarly, when the estimation units 12c and 12d are selected, the threshold values TH ₁ and the threshold values TH 2, ₁ are used, and the threshold values TH 2, ₀ are not particularly used.

By the way, when the estimation units 12a, 12b, 12c, and 12d are assigned numerical values 0 to 3 in decimal, respectively, and when these are converted into binary numbers, the estimation unit 12a is (00) and the estimation unit 12b is (01). ), The estimation unit 12c is (10), and the estimation unit 12d is (11).

When the digits in this binary number are defined as the first digit and the second digit from the upper order, when the first digit is 0, the threshold value TH _2,0 is advanced, and when the first digit is 1, the digit is advanced to the threshold value TH 2,0. This is the case of advancing to the threshold value TH _2,1 . That is, _{k at the threshold values TH 2 and k} corresponds to the numerical value of the first digit. The numerical values of the second digit are the estimation unit 12a (00) or the estimation unit 12b (01) separated by the threshold values TH _2,0 and TH _2,1 and the estimation unit 12c (10) or the estimation unit 12d (11). Corresponds to.

The search unit 14 described above may be realized based on, for example, any reinforcement learning technique. Also in this case, the prediction error PE is used as a parameter used for learning, and the superior estimation unit is searched for.

<Conversion unit 15>
The conversion unit 15 converts the first signal strength TS1 into an optical signal L, for example, as shown in FIG. In this case, the plurality of estimation units 12 generate the prediction sequence PW based on the optical signal L output from the conversion unit 15.

The conversion unit 15 converts the signal intensity (for example, the first signal intensity TS1) into the optical signal L by using, for example, a semiconductor laser. The conversion unit 15 divides the output of the optical signal L output from the semiconductor laser in the feedback loop at regular intervals, and acquires the response output as a virtual node state VN. In this case, the plurality of estimation units 12 acquire different weight information in advance, calculate the weight information in the virtual node state VN, and generate the prediction series PW. The weight information is constructed to estimate the characteristic signal strength of one type of the irregular signal TS.

The conversion unit 15 may be realized by using, for example, any optical reservoir computing technique. Also in this case, the signal strength (for example, the first signal strength TS1) is used as an input parameter, and each estimation unit 12 generates different prediction series PWs.

(Model selection method)
Next, a model selection method to which the present invention is applied will be described with reference to FIG. 7. FIG. 7 is a flowchart showing an example of a model selection method to which the present invention is applied.

<Acquisition step S110>
First, the irregular signal TS is acquired (acquisition step S110). For example, the acquisition unit 11 acquires the irregular signal TS. The acquisition unit 11 outputs the first signal strength TS1 of the irregular signal TS at the first timing, and outputs the second signal strength TS2 of the irregular signal TS at the second timing. The frequency and timing at which the acquisition unit 11 acquires the irregular signal strength TS, and the frequency and timing at which the first signal strength TS1 and the second signal strength TS2 are output are arbitrary.

<Conversion step S150>
Next, for example, the first signal strength TS1 may be converted into an optical signal L (conversion step S150). For example, the conversion unit 15 acquires the first signal intensity TS1 and outputs an optical signal L using a semiconductor laser. The conversion unit 15 acquires the virtual node state VN from the output of the optical signal L.

<Estimation step S120>
Next, a plurality of prediction series PWs are generated (estimation step S120). For example, the plurality of estimation units 12 generate a prediction sequence PW that estimates the signal strength of the irregular signal TS at the second timing based on the first signal strength TS1. Each estimation unit 12 acquires the first signal strength TS1 from the acquisition unit 11 to generate the prediction series PW, and for example, the conversion unit 15 calculates weight information in the virtual node state VN to generate the prediction series PW. May be good. In any case, each of the prediction sequence PWs is suitable for estimating the characteristic signal strength of one type of the irregular signal TS, and different prediction sequence PWs are generated.

<Evaluation step S130>
Next, the prediction error PE is calculated (evaluation step S130). For example, the evaluation unit 13 acquires the second signal strength TS2 of the irregular signal TS at the second timing from the acquisition unit 11. The evaluation unit 13 calculates each prediction error PE of the plurality of prediction series PW with respect to the second signal strength TS2.

<Search step S140>
Next, the superiority estimation unit is searched (search step S140). For example, the search unit 14 acquires a plurality of prediction error PEs from the evaluation unit 13. The search unit 14 searches for a superior estimation unit suitable for estimating the second signal strength TS2 among the plurality of estimation units 12 that have generated the prediction series PW based on the plurality of prediction error PEs.

In the search step S140, for example, the threshold adjustment unit 14a, the signal transmission unit 14b, the comparison unit 14c, and the selection unit 14d may be used to search for the superior estimation unit. In this case, the threshold value adjusting unit 14a adjusts the threshold value TH for selecting the superior estimation unit based on the plurality of prediction error PEs. The signal transmitting unit 14b transmits an irregular selection signal. The comparison unit 14c compares the signal strength of the irregular selection signal with the threshold value TH. The selection unit 14d selects the superiority estimation unit based on the comparison result by the comparison unit 14c.

This completes the model selection method to which the present invention is applied. After performing the search step S140, the acquisition step S110 or the like may be performed again depending on the situation.

According to the model selection device 1 to which the present invention is applied, the search unit 14 searches for an estimation unit 12 (dominant estimation unit) suitable for estimating the second signal strength TS2 based on a plurality of prediction error PEs. Further, each of the plurality of estimation units 12 is optimized for estimating the characteristic signal strength of one type of the irregular signal TS. Therefore, the estimation unit 12 (dominant estimation unit) most suitable for the prediction performance can be searched regardless of the type and timing of the irregular signal TS. This makes it possible to use it to solve problems in various realistic events.

Further, according to the model selection device 1 to which the present invention is applied, a plurality of estimation units 12 formed so as to maximize the performance for a specific aspect are used, and the estimation unit 12 (model) most suitable for the actual aspect is used. An approach of choosing can be realized. This does not require enormous learning or acquisition of generalization ability, and it is possible to achieve both short latency and high throughput.

Further, according to the model selection device 1 to which the present invention is applied, the search unit 14 may have a threshold value adjustment unit 14a, a signal transmission unit 14b, a comparison unit 14c, and a selection unit 14d. Therefore, when searching for the superiority estimation unit, reinforcement learning using optical chaos can be used. This makes it possible to shorten the search time of the superiority estimation unit.

In particular, the threshold value adjusting unit 14a may adjust the threshold value TH for selecting the superior estimation unit based on a plurality of prediction error PEs. Therefore, even when there is a large variation in the prediction error PE, it is possible to carry out the selection following the selection tendency up to the immediately preceding, and it is possible to reduce the selection error due to the variation in the prediction error PE. This makes it possible to suppress a decrease in the accuracy of selecting the superiority estimation unit for the irregular signal TS that changes irregularly.

Further, according to the model selection device 1 to which the present invention is applied, the conversion unit 15 may convert the first signal intensity TS1 into an optical signal L. Therefore, the time for generating the predicted sequence PW can be significantly shortened as compared with the case where the first signal strength TS1 is converted into an object other than the optical signal L. This makes it possible to search for the superior estimation unit most suitable for the prediction performance in a short time.

Further, according to the model selection device 1 to which the present invention is applied, the plurality of estimation units 12 generate a prediction sequence PW based on the optical signal L. Therefore, it is possible to easily realize a combination of the conversion unit 15 that can generate the prediction sequence PW in a short time and the search unit 14 that uses reinforcement learning. This makes it possible to significantly shorten the search time of the superiority estimation unit.

Further, according to the model selection device 1 to which the present invention is applied, the weight information may be constructed to estimate the characteristic signal strength of one type of the irregular signal TS. Therefore, it can be easily constructed as compared with the case of constructing one weight information for estimating a plurality of characteristic signal strengths. This makes it possible to simplify advance preparation.

Further, according to the model selection device 1 to which the present invention is applied, the number of estimation units 12 may be greater than or equal to the number of characteristic signal strength types. Therefore, it is possible to search for all the characteristic signal intensities included in the irregular signal TS under the condition that the estimation unit 12 most suitable for the prediction performance is always included. This makes it possible to improve the search accuracy.

According to the model selection method to which the present invention is applied, the search step S140 searches for an estimation unit 12 (dominant estimation unit) suitable for estimating the second signal strength TS2 based on a plurality of prediction error PEs. Further, each of the plurality of prediction series PWs is generated based on the characteristic signal strength of one of the irregular signal TSs suitable for estimation. Therefore, the estimation unit 12 (dominant estimation unit) most suitable for the prediction performance can be searched regardless of the type and timing of the irregular signal TS. This makes it possible to use it to solve problems in various realistic events.

Next, the results of the simulation performed to confirm the effect of the present invention will be described. In this simulation, based on the prediction series PW that estimates the signal strength of the irregular signal TS at each timing (hereinafter referred to as step), the adjustment transition of the threshold value TH set in the search unit 14 and the estimation as the superior estimation unit. The ratio of correctly selecting the suitable estimation unit 12 was verified.

In this embodiment, the acquisition unit 11 shown in FIG. 6, the conversion unit 15, the two estimation units 12a and 12b, the evaluation unit 13, and the search unit 14 are used, and the search unit 14 is shown in FIG. The threshold adjustment unit 14a, the signal transmission unit 14b, the comparison unit 14c, and the selection unit 14d were used.

In this embodiment, the prediction error PEs of the prediction series PW1 and PW2 generated from the estimation units 12a and 12b are calculated, and the threshold value is selected so that the smaller absolute value of the prediction error PE is preferentially selected. It was set to adjust TH. The threshold value TH was set to be adjusted based on the immediately preceding value. This was set in order to follow the variation in the prediction error PE and prevent a selection error due to a large change in the threshold value TH.

FIG. 8A is a graph showing the prepared irregular signal TS. The horizontal axis of the figure shows the steps, and the vertical axis shows the signal strength at each step. The characteristic signal intensities included in the irregular signal TS are formed based on the Lorentz type time series (range L in FIG. 8A) and the wrestler type time series (range R in FIG. 8A). The signal strength was used. Each signal strength was assumed to change irregularly. In FIG. 8A, the Lorentz type time series changes to the wrestler type time series around 450 steps. Note that FIG. 8A shows the signal strength from 0 step to 1000 step, but in this embodiment, it was carried out up to 4000 step.

In this embodiment, the prediction series PW1 and PW2 are generated by using two estimation units 12a and 12b having weight information that has been constructed (learned) in advance. As the weight information of the estimation unit 12a, the information optimized for estimating the signal strength formed based on the wrestler type time series was used. As the weight information of the estimation unit 12b, the information optimized for estimating the signal strength formed based on the Lorentz type time series was used.

The prediction series PW1 and PW2 generated by the estimation units 12a and 12b based on the above conditions are shown in the graphs of FIGS. 8 (b) and 8 (c). FIG. 8B shows a prediction sequence PW1 estimated by using the estimation unit 12a with respect to the signal strength of the irregular signal TS shown in FIG. 8A. FIG. 8C shows a prediction sequence PW2 estimated by using the estimation unit 12b with respect to the signal strength of the irregular signal TS shown in FIG. 8A.

FIG. 9 is a graph showing the prediction errors PE1 and PE2 of the prediction series PW1 and PW2 with respect to the signal strength of the irregular signal TS in each step. The horizontal axis of FIG. 9 shows the steps, and the vertical axis shows the absolute value of the prediction error PE in each step. The smaller the value of the prediction error PE, the higher the estimation accuracy.

In FIG. 9A, it was confirmed that the value of the prediction error PE after the 450 to 500 steps was smaller than the value of the prediction error PE of the steps before that. This is because the weight information of the estimation unit 12a is optimized for estimating the signal strength formed based on the wrestler type time series. Further, in FIG. 9B, it was confirmed that the value of the prediction error PE before 450 to 500 steps was smaller than the value of the prediction error PE of the subsequent steps. This is because the weight information of the estimation unit 12b used is optimized for estimating the signal strength formed based on the Lorentz type time series.

FIG. 10A is a graph showing the transition of the threshold value TH set in the search unit 14. The horizontal axis of FIG. 10A indicates a step, and the vertical axis indicates a threshold value TH. FIG. 10A shows that the estimation unit 12b is more likely to be selected as the threshold TH increases from 0, and the estimation unit 12a is more likely to be selected as the threshold TH decreases from 0. Show that it will be.

FIG. 10A shows a tendency that the threshold value TH increases from 0 step to about 400 steps. This is because the value of the prediction error PE2 corresponding to the estimation unit 12b is smaller than the value of the prediction error PE1 corresponding to the estimation unit 12a, and the threshold value TH changes so that the estimation unit 12b is preferentially selected as the superior estimation unit. It shows that it was done.

In addition, the threshold TH tends to decrease from 400 steps to a little less than 2000 steps. This is because the value of the prediction error PE1 corresponding to the estimation unit 12a is smaller than the value of the prediction error PE2 corresponding to the estimation unit 12b, and the threshold value TH changes so that the estimation unit 12a is preferentially selected as the superior estimation unit. It shows that it was done.

Further, the threshold value TH tends to increase from a little less than 2000 steps to a little less than 3000 steps. This is because the value of the prediction error PE2 corresponding to the estimation unit 12b is smaller than the value of the prediction error PE1 corresponding to the estimation unit 12a, and the threshold value TH changes so that the estimation unit 12b is preferentially selected as the superior estimation unit. It shows that it was done.

Further, the threshold value TH tends to decrease from a little less than 3000 steps to about 4000 steps. This is because the value of the prediction error PE1 corresponding to the estimation unit 12a is smaller than the value of the prediction error PE2 corresponding to the estimation unit 12b, and the threshold value TH changes so that the estimation unit 12a is preferentially selected as the superior estimation unit. It shows that it was done.

FIG. 10B is a graph showing the ratio of correctly selecting the estimation unit 12 suitable for estimation as the superior estimation unit. The horizontal axis of FIG. 10B shows the steps, and the vertical axis shows the correctly selected ratios (minimum 0.0, maximum 1.0).

In FIG. 10B, the correctly selected ratio is maintained at 1.0 from 0 step to about 400 steps. That is, it was confirmed that the selection of the estimation unit 12b as the superior estimation unit was correctly executed.

On the other hand, after 400 steps, the correctly selected ratio suddenly changes from 1.0 to around 0.0. This is due to the change from the Lorentz type time series to the signal strength formed based on the wrestler type time series. That is, it means that the threshold value TH for giving priority to the estimation unit 12b does not follow a sudden change in signal strength.

However, the percentage selected correctly increased to about 0.9 in about 1000 steps and maintained at 1.0 after 1400 steps. That is, it was confirmed that the threshold value TH was adjusted each time the steps were repeated, and the estimation unit 12a was selected as the superior estimation unit.

Further, after a little less than 2000 steps, the correctly selected ratio suddenly changes from 1.0 to around 0.0. This is due to the change from the wrestler type time series to the signal strength formed based on the Lorentz type time series. That is, it means that the threshold value TH for giving priority to the estimation unit 12a does not follow a sudden change in signal strength.

However, the proportion correctly selected increased to about 1.0 in about 2200 steps and maintained at 1.0 after 2200 steps. That is, it was confirmed that the threshold value TH was adjusted each time the steps were repeated, and the estimation unit 12b was selected as the superior estimation unit.

Further, after a little less than 3000 steps, the correctly selected ratio suddenly changes from 1.0 to around 0.0. This is due to the change from the Lorentz type time series to the signal strength formed based on the wrestler type time series. That is, it means that the threshold value TH for giving priority to the estimation unit 12b does not follow a sudden change in signal strength.

However, the proportion correctly selected increased to about 1.0 at about 3200 steps and maintained at 1.0 after 3200 steps. That is, it was confirmed that the threshold value TH was adjusted each time the steps were repeated, and the estimation unit 12a was selected as the superior estimation unit.

From the above, it was confirmed that the estimation unit 12 (dominant estimation unit) most suitable for the prediction performance can be searched for the irregular signal TS whose characteristic signal strength changes irregularly. This makes it possible to use it to solve problems in various realistic events.

The acquisition unit 11, estimation unit 12, evaluation unit 13, search unit 14, and conversion unit 15 described above can be realized by using, for example, a semiconductor integrated circuit (LSI), or, for example, a computer having a known microprocessor or memory. May be realized by executing the program.

Although some embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These novel embodiments can be implemented in various other embodiments, and various omissions, replacements, and changes can be made without departing from the gist of the invention. These embodiments and modifications thereof are included in the scope and gist of the invention, and are also included in the scope of the invention described in the claims and the equivalent scope thereof.

1: Model selection device 11: Acquisition unit 12: Estimating unit 13: Evaluation unit 14: Search unit 14a: Threshold adjustment unit 14b: Signal transmission unit 14c: Comparison unit 14d: Selection unit 15: Conversion unit 4: Decision unit 91: Semiconductor laser 92: Coupler 93: Attenuator 94: Reflector 95: Optical isolator 96: Attenuator 97: Light receiving unit 98: Oscillator L: Optical signal PE: Prediction error PW: Prediction series S110: Acquisition step S120: Estimation step S130: Evaluation step S140: Search step S150: Conversion step TH: Threshold TS: Irregular signal TS1: First signal strength TS2: Second signal strength

Claims (6)

An acquisition unit that acquires an irregular signal in which at least two types of characteristic signal strengths change irregularly,
A plurality of estimation units that generate a prediction series that estimates the signal strength of the irregular signal at the second timing after the first timing based on the first signal strength of the irregular signal at the first timing.
An evaluation unit that acquires the second signal strength of the irregular signal at the second timing and calculates the prediction error of each of the plurality of prediction series with respect to the second signal strength.
A search unit for searching a superior estimation unit suitable for estimating the second signal strength among the plurality of estimation units based on the plurality of prediction errors is provided.
A model selection device, wherein each of the plurality of estimation units is optimized for estimating the characteristic signal strength of one type of the irregular signal. The search unit
A threshold adjustment unit that adjusts a threshold value for selecting the superior estimation unit based on the plurality of prediction errors, and a threshold adjustment unit.
A signal transmitter that emits a selection irregular signal whose signal strength changes irregularly,
A comparison unit that compares the signal strength of the irregular signal for selection with the threshold value,
The model selection device according to claim 1, further comprising a selection unit for selecting the superiority estimation unit based on the comparison result by the comparison unit. Further, a conversion unit for converting the first signal strength into an optical signal is provided.
The model selection device according to claim 1 or 2, wherein the plurality of estimation units generate the prediction series based on an optical signal output from the conversion unit. The plurality of estimation units acquire different weight information in advance, calculate the weight information in the virtual node state acquired from the output of the optical signal, and generate the prediction series.
The model selection device according to claim 3, wherein the weight information is constructed to estimate the characteristic signal strength of one of the irregular signals. The model selection device according to any one of claims 1 to 4, wherein the number of the plurality of estimation units is equal to or greater than the number of the characteristic signal strength types. The acquisition step of acquiring an irregular signal in which at least two kinds of characteristic signal strengths change irregularly, and
Based on the first signal strength of the irregular signal at the first timing, an estimation step of generating a plurality of prediction sequences for estimating the signal strength of the irregular signal at the second timing after the first timing, and an estimation step.
An evaluation step of acquiring the second signal strength of the irregular signal at the second timing and calculating the prediction error of each of the plurality of prediction series with respect to the second signal strength.
A search step for searching for a superior estimation unit suitable for estimating the second signal strength among a plurality of estimation units that have generated the prediction series based on the plurality of prediction errors is provided.
A model selection method, characterized in that each of the plurality of prediction sequences is generated based on one of the irregular signals suitable for estimating the characteristic signal strength.

Images